Heng, Xin · Sep 19, 2026
Xin Heng argues agent workflows can skip ready steps a request does not need
Markdown · JSON · Original source
In the abstract for LazyAgent, Xin Heng argues that agent runtimes often run every step as soon as it is ready, even when a request does not need that step. LazyAgent runs a ready step only when the current request needs it, and keeps that answer exact as the plan changes. Heng reports that, across 303 requests on 154 production workflows, an eager runtime that stops when the goal is met spends 62.2% of its on work the request never needed. In the reported comparisons, LazyAgent saves , tokens, or dollars in several settings, while tying when a request needs the whole graph. The abstract also reports that projection saves 80-89% of a shared CI step, where the identical eager control saves 0.0%. The central argument is that knowing which steps are demanded can avoid unnecessary work and help choose among candidate workflows. Heng says public agent are designed like eager runtimes and therefore cannot show this difference; the paper calls on industry to build benchmarks that measure workflow efficiency. The abstract reports results but does not describe the experimental methods in detail. What benchmark design would measure workflow efficiency while preserving matched quality?
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
cs.AI · Article
Source details
Original source date: Sep 19, 2026 (2026-09-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T01:39:38.952Z. First observed: 2026-11-02T05:03:16.428Z. Retrieved: 2026-11-02T09:48:33.239Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
These are results reported in the paper's abstract; the supplied text does not provide experimental methods or enough detail to assess how the comparisons were conducted. The abstract reports no losses in its stated comparisons, but does not establish performance beyond the tested settings.
Open question
What design would measure workflow efficiency while preserving matched quality?
Source-stated
Xin Heng argues that eager agent runtimes may execute steps that are ready but unnecessary for a request.
Neither these runtimes nor their benchmarks distinguish a step that is ready from a step the request needs
Source-stated
Across the reported production workflows, an eager runtime that stops once the goal is met spent 62.2% of its CPU on work the request never needed.
across 303 requests on 154 production workflows, an eager runtime that stops once the goal is met spends 62.2% of its CPU on work the request never needed.
Source-stated
LazyAgent runs a ready step only when the current request needs it, and the abstract says its demanded set is derived in one traversal with no model calls.
run a ready step only if the current request needs it, and keep that answer exact as the plan changes. A demanded set derived from the program decides, in one traversal with no model calls
Source-stated
The abstract reports savings across several settings, including time, tokens, and dollars, with no losses in the stated comparisons.
Measured live against that eager runtime, LazyAgent saves 51.7% and 52.1% of container time on release gates where a frontier LLM writes the patches, 12.8% of tokens on DataSciBench plans at matched quality, and 24.0% of dollars inside a live LangGraph application, with no losses.
Source-stated
The paper says public agent benchmarks cannot show the difference because they are designed like eager runtimes, and calls for benchmarks that measure workflow efficiency.
Public agent benchmarks are designed the same way as eager runtimes, so they cannot show this difference. We compose our test programs from their tasks and call on industry to build benchmarks that measure how efficiently agent workflows run.
Source and contributor context
Heng, Xin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
CPU
The central processing unit: a general-purpose processor that carries out a computer's instructions.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 65f78b03-9afc-4a24-8295-2ffbbf97f09e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T01:39:38.952Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5e510b611a32b2d45e12e3b7ce88bbb19adad1219dbfba2ce64de94fa00b044a. Source text hash: 1a10926950d7b74942ae983c95f7cea74f39834c01d8f55132370aea6677c5e1.
Revision and source-check history
Version 1 · 2026-11-11T01:39:38.952Z
Xin Heng argues agent workflows can skip ready steps a request does not need
In the abstract for LazyAgent, Xin Heng argues that agent runtimes often run every step as soon as it is ready, even when a request does not need that step. LazyAgent runs a ready step only when the current request needs it, and keeps that answer exact as the plan changes. Heng reports that, across 303 requests on 154 production workflows, an eager runtime that stops when the goal is met spends 62.2% of its on work the request never needed. In the reported comparisons, LazyAgent saves , tokens, or dollars in several settings, while tying when a request needs the whole graph. The abstract also reports that projection saves 80-89% of a shared CI step, where the identical eager control saves 0.0%. The central argument is that knowing which steps are demanded can avoid unnecessary work and help choose among candidate workflows. Heng says public agent are designed like eager runtimes and therefore cannot show this difference; the paper calls on industry to build benchmarks that measure workflow efficiency. The abstract reports results but does not describe the experimental methods in detail. What benchmark design would measure workflow efficiency while preserving matched quality?
Original source
cs.AI · Article
Source details
Original source date: Sep 19, 2026 (2026-09-19).
Date basis: Date basis not provided.
Added to Escape Meridian: 2026-11-11T01:39:38.952Z. First observed: 2026-11-02T05:03:16.428Z. Retrieved: 2026-11-02T09:48:33.239Z.
Limits
These are results reported in the paper's abstract; the supplied text does not provide experimental methods or enough detail to assess how the comparisons were conducted. The abstract reports no losses in its stated comparisons, but does not establish performance beyond the tested settings.
Open question
What design would measure workflow efficiency while preserving matched quality?
Source-stated
Xin Heng argues that eager agent runtimes may execute steps that are ready but unnecessary for a request.
Neither these runtimes nor their benchmarks distinguish a step that is ready from a step the request needs
Source-stated
Across the reported production workflows, an eager runtime that stops once the goal is met spent 62.2% of its CPU on work the request never needed.
across 303 requests on 154 production workflows, an eager runtime that stops once the goal is met spends 62.2% of its CPU on work the request never needed.
Source-stated
LazyAgent runs a ready step only when the current request needs it, and the abstract says its demanded set is derived in one traversal with no model calls.
run a ready step only if the current request needs it, and keep that answer exact as the plan changes. A demanded set derived from the program decides, in one traversal with no model calls
Source-stated
The abstract reports savings across several settings, including time, tokens, and dollars, with no losses in the stated comparisons.
Measured live against that eager runtime, LazyAgent saves 51.7% and 52.1% of container time on release gates where a frontier LLM writes the patches, 12.8% of tokens on DataSciBench plans at matched quality, and 24.0% of dollars inside a live LangGraph application, with no losses.
Source-stated
The paper says public agent benchmarks cannot show the difference because they are designed like eager runtimes, and calls for benchmarks that measure workflow efficiency.
Public agent benchmarks are designed the same way as eager runtimes, so they cannot show this difference. We compose our test programs from their tasks and call on industry to build benchmarks that measure how efficiently agent workflows run.
Source and contributor context
Heng, Xin · Article author
These are credits stated by the source; they do not establish a profile match or authorship of every passage.
Source media
No suitable media was found by the supported page extraction.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
CPU
The central processing unit: a general-purpose processor that carries out a computer's instructions.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
arXiv
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
TIME
TIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Editorial provenance
Version 1. Publication ID: 65f78b03-9afc-4a24-8295-2ffbbf97f09e.
Prepared by gpt-6-luna; grounding review by gpt-6-luna.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-11T01:39:38.952Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: material_development.
Observation: 5e510b611a32b2d45e12e3b7ce88bbb19adad1219dbfba2ce64de94fa00b044a. Source text hash: 1a10926950d7b74942ae983c95f7cea74f39834c01d8f55132370aea6677c5e1.
uncertain · 2026-11-11T01:39:38.952Z · anthropic/claude-opus-5.5
Publication: 65f78b03-9afc-4a24-8295-2ffbbf97f09e
attribution — pass: The page lists Xin Heng as the sole author. The draft attributes the argument and results to Heng and the abstract, and frames them as reported claims, not established facts.
Authors:Xin Heng
summary — uncertain: Two problems. (1) Scope: the source limits the claim to runtimes 'that plan before acting'. The draft drops that qualifier and adds 'often'. Smallest repair: restore 'agent runtimes that plan before acting'. (2) Readability: central terms go unexplained for a general reader. These include agent runtime, eager runtime, tokens, the whole graph, demanded set, and 'projection saves 80-89% of a shared CI step'. Briefly explaining eager versus demand-driven execution, and either glossing or dropping the projection/CI line, would fix this. The question is also duplicated at the end of the summary.
Agent runtimes that plan before acting run every step as soon as it is ready.
date — pass: originalPublished is 2026-09-19, which matches the v1 submission date. The 30 Oct 2026 revision and the November feed date are correctly not used.
Submitted on 19 Sep 2026 (v1), last revised 30 Oct 2026 (this version, v2)
support — pass: Each claim matches the abstract verbatim: the 62.2% CPU figure, the rule and its single traversal, the reported savings with no losses, and the benchmark critique. The limitations correctly note that only abstract-level detail is available. The question about matched quality follows from the source's 'at matched quality' wording and is not a leading claim.
12.8% of tokens on DataSciBench plans at matched quality, and 24.0% of dollars inside a live LangGraph application, with no losses.
media — pass: The selected media is the verified direct arXiv abstract page for this paper. It is the most direct access to the source.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe central processing unit: a general-purpose processor that carries out a computer's instructions.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivThe central processing unit: a general-purpose processor that carries out a computer's instructions.
Definition referenceTIME is a news magazine publishing reporting on world affairs, business, technology and the workers behind digital services.
Explore TIMEA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referencearXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.
Explore arXivA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference