Brenda Potts · Nov 16, 2026
Microsoft Research presents TokenEconomizer, which learns to reduce coding-agent costs
Markdown · JSON · Original source
presents TokenEconomizer, an open-source plugin that learns from completed coding-agent sessions in a repository. The researchers argue that agents waste tokens rediscovering familiar facts, loading unused tools, and repeating failed approaches. The plugin uses session logs to build repository-specific “micro-skills,” prune tools, and detect when an agent is stuck; it can suggest another approach without modifying the agent or its underlying model. Removed tools can be restored, and if an agent remains stuck after a suggestion, TokenEconomizer alerts the developer. In evaluations on 277 SWE-Bench tasks from six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%, without performance loss in the tested repository-combination pairs. Early tasks can cost more while the plugin learns, but cumulative costs fell below the default in every repository by the sixth task. The reported savings include TokenEconomizer’s own model calls. Researchers also report lower costs on sequential web-development and security-auditing tasks. They say lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T19:47:20.024Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T19:46:24.952Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 5. Publication ID: d995de36-d69f-43ac-8cd9-86b4b5ae9df4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T19:47:20.024Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2128ecfa9bb98f6cfaa60cae2fdf1806e873098f3831cf86b5b40815a5dd6f6b. Source text hash: 85f9453bf1206984ea028f15d00f77ff90858fada28765c7cda01bde44042098.
Correction reason: The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Revision and source-check history
Version 5 · 2026-11-16T19:47:20.024Z · The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Microsoft Research presents TokenEconomizer, which learns to reduce coding-agent costs
presents TokenEconomizer, an open-source plugin that learns from completed coding-agent sessions in a repository. The researchers argue that agents waste tokens rediscovering familiar facts, loading unused tools, and repeating failed approaches. The plugin uses session logs to build repository-specific “micro-skills,” prune tools, and detect when an agent is stuck; it can suggest another approach without modifying the agent or its underlying model. Removed tools can be restored, and if an agent remains stuck after a suggestion, TokenEconomizer alerts the developer. In evaluations on 277 SWE-Bench tasks from six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%, without performance loss in the tested repository-combination pairs. Early tasks can cost more while the plugin learns, but cumulative costs fell below the default in every repository by the sixth task. The reported savings include TokenEconomizer’s own model calls. Researchers also report lower costs on sequential web-development and security-auditing tasks. They say lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T19:47:20.024Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T19:46:24.952Z.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 5. Publication ID: d995de36-d69f-43ac-8cd9-86b4b5ae9df4.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T19:47:20.024Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 2128ecfa9bb98f6cfaa60cae2fdf1806e873098f3831cf86b5b40815a5dd6f6b. Source text hash: 85f9453bf1206984ea028f15d00f77ff90858fada28765c7cda01bde44042098.
Correction reason: The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Version 4 · 2026-11-16T19:04:17.117Z · The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Microsoft Research presents TokenEconomizer, which learns to reduce coding-agent costs
presents TokenEconomizer, an open-source plugin that learns from completed coding-agent sessions in a repository. The researchers’ premise is that agents can waste tokens rediscovering repository facts, loading tools they do not use, and repeating failed approaches. TokenEconomizer turns session logs into repository-specific “micro-skills,” prunes tools the agent is unlikely to need, and uses a judge model to detect when an agent is stuck and suggest another approach. It works without modifying the agent or its underlying model. Removed tools can be restored; if an agent remains stuck after a suggestion, the plugin alerts the developer. In evaluations on 277 SWE-Bench tasks from six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%, without performance loss in the tested repository-combination pairs. Early tasks can cost more while the plugin learns, but cumulative costs fell below the default in every repository by the sixth task. Reported savings include the plugin’s own model calls. The researchers also report lower costs on sequential web-development and security-auditing tasks. They say lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T19:04:17.117Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T19:02:44.005Z.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 4. Publication ID: bc1db08d-9e03-4bb9-88e0-23b26bd75552.
Prepared by gpt-6-luna; stream-sentence-removal-v1: removed claim 6; grounding review by gpt-6-luna.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-16T19:04:17.117Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: e9ff3bb4067a49fdbf6ffc7c02d9b1f8a92c0ff787ff507036981389341af706. Source text hash: 0802273a3be4df59b02d613e401b77b3c89ef1d2b76325fdd184abb125659dd8.
Correction reason: The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Version 3 · 2026-11-16T18:19:23.243Z · The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Microsoft Research presents TokenEconomizer, which learns to cut coding-agent costs
presents TokenEconomizer, an open-source plugin that learns from completed coding-agent sessions in a repository. Its premise is that agents often repeat work across sessions: rediscovering facts, loading unused tools and revisiting failures. The plugin adds repository-specific “micro-skills” after repeated evidence, prunes tools history suggests will go unused, and uses a judge model to identify failure loops and suggest a different approach. It works without modifying the agent or its underlying model; removed tools can be restored, and if the agent remains stuck after a suggestion, the plugin alerts the developer. In evaluations on 277 SWE-Bench tasks from six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%, without performance loss in the tested repository-combination pairs. Early tasks could cost more while the plugin learns, but cumulative costs fell below the default in every repository by the sixth task. Reported savings include the plugin’s own model calls. The researchers also report lower costs on sequential web-development and security-auditing tasks. They say lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T18:19:23.243Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T18:18:07.904Z.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Escape Meridian interpretation
The results suggest that repository-specific experience can reduce token use without reducing model reasoning effort; the researchers report that lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
By contrast, lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 3. Publication ID: b768e55d-2685-413f-8c98-487d0bddfdba.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T18:19:23.243Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 426a136326fe74c2f54beaf8ee4925c619d7c5a2b67b0d434abcde7f79cd77a1. Source text hash: d47030acdb0819f4a749fb53d26ebc482404aa7571a298ca1649adec222a50b8.
Correction reason: The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Version 2 · 2026-11-16T18:03:37.632Z · The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Microsoft Research says TokenEconomizer learns from coding sessions to reduce costs
describes TokenEconomizer, an open-source plugin that learns from completed AI coding-agent sessions in a repository to reduce waste in later sessions, without changing the agent or its model. It adds repository-specific “micro-skills” after repeated evidence, prunes tools history suggests will go unused, and intervenes when a judge model identifies a failure loop. Removed tools can be restored; if the agent remains stuck after a suggested change, the plugin alerts the developer. In evaluations spanning 277 SWE-Bench tasks, six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%, without performance loss in the tested repository-combination pairs. Cumulative costs fell below the default in every repository by the sixth task, though early tasks could cost more while the plugin learned. The reported savings include the plugin’s own model calls. The researchers also report lower costs on sequential web-development and security-auditing tasks. Escape Meridian interpretation: the comparison suggests that reducing reasoning effort is not equivalent to learning repository-specific ways to save costs; the researchers say reduced reasoning effort saved similar or larger amounts but substantially degraded performance.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T18:03:37.632Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T18:02:10.275Z.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Escape Meridian interpretation
The results suggest that repository-specific experience can reduce token use without reducing model reasoning effort; the researchers report that lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
By contrast, lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 2. Publication ID: 3262db45-ae3a-46e8-8aad-1d3f54749f45.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T18:03:37.632Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 6bc1818e245388399e53c619a021397abe3ebd2bb81917c282f701880ca16cc8. Source text hash: 328cc46dccc4b09b21c9c398359542c67f73751230644778f36fc9c9f15d3792.
Correction reason: The summary was revised to foreground how the plugin learns and intervenes, retain the comparison with reduced reasoning effort, and clarify that reported results come from evaluated settings.
Version 1 · 2026-11-16T17:17:22.218Z
Microsoft Research presents TokenEconomizer, which learns from coding sessions to cut costs
describes TokenEconomizer, an open-source plugin that uses an AI coding agent’s past work in a repository to reduce repeated effort in later sessions. It does not modify the agent or its underlying model. The plugin records repository-specific patterns as “micro-skills,” removes tools history suggests are unlikely to be used, and intervenes when a judge model identifies a failure loop. Developers can restore removed tools; if an agent remains stuck after a suggested change in approach, the plugin alerts them. In tests on 277 SWE-Bench tasks across six repositories and five agent–model combinations, the researchers report average cost reductions of 18.4–30.9%. They say costs fell in every tested repository-combination pair without performance loss, while Pass@1 rose in four combinations. Savings include TokenEconomizer’s own model calls. Early tasks can cost more as the plugin learns, but cumulative costs fell below the default by the sixth task in every repository. The researchers also report lower costs on sequential web-development tasks and security-auditing tasks, including tasks in new repositories. They say lowering the model’s reasoning effort saved similar or larger amounts but substantially degraded performance.
Original source
· Article
Source details
Original source date: Nov 16, 2026 (2026-11-16T17:01:35+00:00).
Date basis: Date stated by the publisher.
Added to Escape Meridian: 2026-11-16T17:17:22.218Z. First observed: 2026-11-16T17:03:40.126Z. Retrieved: 2026-11-16T17:15:52.808Z.
Limits
The reported evaluation covers 277 SWE-Bench tasks from six repositories and five agent–model combinations; results beyond those tested settings are not established here. The source reports cost and performance outcomes from evaluations, not a guarantee of savings or task success in other use. The source says early tasks can cost more while TokenEconomizer learns, before cumulative costs fall below the default.
Open question
How would TokenEconomizer’s cost savings and task performance vary across repositories and coding tasks beyond those evaluated?
Source-stated
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository and reduces waste without changing the agent or its underlying model.
TokenEconomizer is an open-source plugin that learns from an agent’s past sessions in a repository to reduce waste in future sessions, without modifying the agent or its underlying model.
Source-stated
The plugin uses micro-skills, tool pruning, and failure-loop intervention to guide later sessions, with safeguards for restoring tools and alerting developers.
Tool pruning removes tools that repository history suggests the agent is unlikely to use, while always retaining core tools for running commands and editing files. Removed tools can be restored if needed.
Source-stated
In evaluations on 277 SWE-Bench tasks, researchers report cost reductions of 18.4–30.9% across five agent–model combinations, with no performance loss in the tested repository-combination pairs.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
Source-stated
The researchers report that cumulative costs fell below the default by the sixth task in every repository, although early tasks could cost more while the plugin learned.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
Source-stated
The researchers report lower costs on sequential web-development and security-auditing tasks as well.
The savings also carried over to sequential web development, with costs 16.0–21.2% lower on Web-Bench. On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
Escape Meridian interpretation
The results suggest that repository-specific experience can reduce token use without reducing model reasoning effort; the researchers report that lowering reasoning effort saved similar or larger amounts but substantially degraded performance.
By contrast, lowering the model’s reasoning effort saved similar or larger amounts but led to substantial performance degradation.
Source and contributor context
Brenda Potts · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
Media declared by the original publisher; availability checked separately from the editorial reading.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Microsoft Research
Primary technical research with named authors, papers and implementation details.
Editorial provenance
Version 1. Publication ID: 793ae8ce-f7f6-4801-895a-19cd19990b74.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-16T17:17:22.218Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 65eb0f2082b9759a22afec59ae678c4376485278237bcc990b89108f64d75a1d. Source text hash: d47030acdb0819f4a749fb53d26ebc482404aa7571a298ca1649adec222a50b8.
uncertain · 2026-11-16T17:17:22.218Z · anthropic/claude-opus-5.5
Publication: 793ae8ce-f7f6-4801-895a-19cd19990b74
attribution — pass: The reading consistently attributes the plugin description and results to Microsoft Research and its researchers, and source findings remain reported claims. The headline names only Microsoft Research, although the plugin is described as coming from UC Berkeley, Microsoft Research, and Microsoft Security. Because the blog is a Microsoft Research publication and the draft does not claim sole authorship, this omission is minor. Adding the collaborators would be a modest fidelity improvement.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security
summary — uncertain: The draft says security-auditing tasks included tasks in new repositories, but the source says every security-auditing task was in a new repository. That detail matters because it shows savings without accumulated repository history. Smallest repair: change it to say every security-auditing task was in a new repository. SWE-Bench and Pass@1 also appear without plain explanation; tooltips or brief glosses would make the evaluation understandable to non-specialists. One possible gloss is that SWE-Bench is a benchmark of real bug fixes in open-source Python projects.
On security auditing, where every task is in a new repository, costs were 6.4–19.6% lower.
date — pass: The evidence gives an original publication date of 2026-11-16 from the publisher page. The feed timestamp matches it, and nothing indicates an earlier first publication.
support — uncertain: The summary says developers can restore removed tools. The source only says, in passive voice, that removed tools can be restored if needed, so it does not establish who restores them; the controller might do it automatically. Smallest repair: say removed tools can be restored if needed. The draft's no-performance-loss wording follows the source's claim. However, the source also says it saved the most, apparently relative to compared approaches, and its at-a-glance section says success rates remained broadly stable. The draft does not contradict either point. The interpretation claim is labeled and supported.
Removed tools can be restored if needed.
media — pass: The selected media is the verified original Microsoft Research blog article. It provides direct, relevant access to the source.
uncertain · 2026-11-16T18:03:37.632Z · anthropic/claude-opus-5.5
Publication: 3262db45-ae3a-46e8-8aad-1d3f54749f45
attribution — pass: The post is published on Microsoft Research's blog, so 'Microsoft Research describes/says' fits, and results are framed as researcher reports. The source names the plugin as a joint UC Berkeley, Microsoft Research and Microsoft Security project. Mentioning UC Berkeley would be more complete, but its absence does not misattribute the work.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security, turns completed sessions into guidance for subsequent ones.
summary — uncertain: The mechanisms are explained clearly, but core context for a non-specialist is missing. 'SWE-Bench' is never described as a benchmark of real bug fixes in open-source Python projects. The draft also never says the 'costs' are tokens consumed by every model interaction, which is the core problem the post addresses. Smallest repair: add 'real bug-fixing tasks from SWE-Bench' and one sentence saying agents pay in tokens for instructions, tool descriptions and history sent on each turn. The question is substantive and non-leading.
SWE-Bench, a benchmark of real bug fixes in open-source Python projects
date — pass: evidence.originalPublished (2026-11-16T17:01:35Z, publisher page) matches the feed entry. Nothing in the text suggests an earlier first publication of this post.
support — uncertain: The draft's 'without performance loss in the tested repository-combination pairs' stretches an ambiguous source sentence. That sentence could instead mean TokenEconomizer saved the most among methods without performance loss. Elsewhere the source says success rates 'remained broadly stable' and Pass@1 rose in four of five combinations, which leaves room for some decline. Smallest repair: say success rates remained broadly stable, with Pass@1 rising in four of five combinations. The reasoning-effort interpretation is labeled as interpretation and sourced. Other figures (18.4–30.9%, sixth task, Web-Bench, security auditing) match the source.
Costs fell in every repository tested, while task success rates remained broadly stable.
media — pass: The selected media is the verified direct link to the original Microsoft Research blog post. That is the most direct way for readers to reach the source.
uncertain · 2026-11-16T18:19:23.243Z · anthropic/claude-opus-5.5
Publication: b768e55d-2685-413f-8c98-487d0bddfdba
attribution — pass: Presenting this as a Microsoft Research work is supported because it appears on the Microsoft Research blog. The source, however, calls it a joint plugin from UC Berkeley, Microsoft Research and Microsoft Security. Smallest repair: add 'with UC Berkeley and Microsoft Security' in the first sentence. Results are attributed to the researchers throughout.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security
summary — uncertain: The reading is faithful and clear about the mechanisms, but the central evidence term SWE-Bench is never explained. The source defines it as a benchmark of real bug fixes in open-source Python projects. 'Tokens', which drive the cost being cut, and 'reasoning effort' are also unexplained. A non-specialist cannot tell what was tested or why sessions cost money. Smallest repair: add a plain-language gloss of SWE-Bench and a note that costs come from tokens sent to the model each turn. The question is a fair open issue.
SWE-Bench, a benchmark of real bug fixes in open-source Python projects
date — pass: originalPublished 2026-11-16T17:01:35Z comes from the publisher page and matches the feed entry. Nothing in the source suggests an earlier first publication of this blog post.
support — uncertain: The draft says there was no performance loss in the tested repository-combination pairs. The source sentence is ambiguous: it says TokenEconomizer saved the most in all 30 pairs, which can be read as a comparison among non-degrading methods. Elsewhere the source says success rates remained 'broadly stable' and Pass@1 rose in four of five combinations. That implies possible small variation in the fifth combination. Smallest repair: say 'task success rates remained broadly stable (Pass@1 rose in four of five combinations)'. Other claims, including the labeled interpretation, are source-grounded.
Costs fell in every repository tested, while task success rates remained broadly stable.
media — pass: The selected media is the verified direct article link to the original Microsoft Research blog post. It is the most direct access to the source.
passed · 2026-11-16T19:04:17.117Z · anthropic/claude-opus-5.5
Publication: bc1db08d-9e03-4bb9-88e0-23b26bd75552
attribution — pass: The reading presents TokenEconomizer as described on the Microsoft Research blog, and it attributes findings to 'the researchers.' The source also credits UC Berkeley and Microsoft Security. Naming them would be more complete, but presenting it as a Microsoft Research presentation is not wrong.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security, turns completed sessions into guidance for subsequent ones.
summary — pass: The summary leads with the actor and the development. It explains the three waste patterns and the three mechanisms in plain language, and it keeps the conditions: early tasks cost more, the plugin's own model calls are counted, and the reduced-reasoning comparison is included. Terms like tokens and SWE-Bench can be handled by tooltips. A small optional repair: noting that security-auditing tasks each used a new repository would clarify the smaller savings there. The question is a genuine open issue about results beyond the tested settings.
Early tasks can cost more while TokenEconomizer learns, but cumulative costs fell below the default in every repository by the sixth task
date — pass: originalPublished is 2026-11-16T17:01:35Z, based on the publisher page. It matches the feed timestamp, and no earlier first publication is indicated.
support — pass: Each claim matches the source text: the plugin's description, the mechanisms and safeguards, the 18.4–30.9% savings on 277 tasks, the sixth-task cumulative result, and the Web-Bench and security-auditing figures. The reading reports no performance loss across the 30 pairs, consistent with the source. The limitations correctly confine results to the evaluated settings without inventing benefits.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
media — pass: The selected media is the verified direct link to the original Microsoft Research blog post. That is the most direct access to the source.
uncertain · 2026-11-16T19:47:20.024Z · anthropic/claude-opus-5.5
Publication: d995de36-d69f-43ac-8cd9-86b4b5ae9df4
attribution — uncertain: The title and lead say 'Microsoft Research presents TokenEconomizer.' The post does appear on the Microsoft Research blog, but the source describes the plugin as a joint work of UC Berkeley, Microsoft Research and Microsoft Security. The authors listed include Berkeley professors. Smallest repair: say 'an open-source plugin from UC Berkeley, Microsoft Research and Microsoft Security' in the lead. Results remain correctly framed as the researchers' reports.
TokenEconomizer, an open-source plugin from UC Berkeley, Microsoft Research, and Microsoft Security, turns completed sessions into guidance for subsequent ones.
summary — uncertain: The reading is faithful, but several central terms are left unexplained for a general reader. 'Micro-skills' are one-line repository notes, SWE-Bench is a benchmark of real bug fixes in Python projects, and 'reasoning effort' and 'tokens' are not explained. The draft also drops the profiling evidence behind the argument: 30% repeated actions, 16 of 23 tools never called, and failed tasks taking 32% of spending. Smallest repair: briefly define micro-skills and SWE-Bench, and restore one or two of those profiling figures.
Micro-skills capture repository-specific knowledge in one-line notes.
date — pass: The originalPublished value of 2026-11-16T17:01:35Z comes from the publisher page, matches the feed entry, and is consistent with first observation shortly afterward. No earlier publication is indicated.
support — pass: Each claim matches a source statement. These include the cost reductions of 18.4–30.9%, 'without performance loss' across 30 pairs, cumulative savings by the sixth task, and the web-development and security-auditing figures. They also include the inclusion of the plugin's own model calls and the reasoning-effort comparison. The limitations and the question stay within the evidence and do not overclaim real-world reliability.
TokenEconomizer lowered the average cost by 18.4–30.9%, as shown in Figure 3, and saved the most in all 30 repository-combination pairs without performance loss.
media — pass: The selected media is the verified link to the original Microsoft Research blog post. It gives readers the most direct access to the actual source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
Primary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft ResearchPrimary technical research with named authors, papers and implementation details.
Explore Microsoft Research