Cohere · Apr 19, 2026
Leon Engländer and co-authors argue in Cohere’s paper that agents ignore useful discoveries
Markdown · JSON · Original source
Cohere’s paper, published April 19, 2026, by Leon Engländer, Sophia Althammer, Ahmet Üstün, Matthias Gallé and Tom Sherborne, argues that current -based agents often fail to act on useful information they discover in their environments. The authors call the missing capability “environmental curiosity”: recognizing and investigating unexpected but relevant observations. Across Terminal-Bench, SWE-Bench and AppWorld, they exposed agents to complete task solutions. Agents found solutions in many Terminal-Bench runs, but used them much less often; in AppWorld, they often saw documentation explicitly identifying a command’s solution, yet rarely exploited it. The authors identify available tools in the , and training data distribution as factors affecting environmental curiosity. They report that configurations maximizing curiosity also perform best on the unmodified , but even jointly optimized agents ignore discovered solutions in most trials. Their argument is that agents use environments to fetch expected information without reliably revising their strategy when useful, unexpected evidence appears. The authors’ results point to a gap between noticing a solution and adapting behavior to use it; how to close that gap across tasks remains unknown.
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
Cohere · Article
Source details
Original source date: Apr 19, 2026 (2026-04-19T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-09T05:53:23.367Z. First observed: 2026-11-09T03:44:12.661Z. Retrieved: 2026-11-09T03:45:36.558Z.
Referenced by
Recorded citations and public file attachments. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on November 9, 2026; originally published April 19, 2026. The supplied evidence is the publisher page’s abstract, not the paper’s full methods or results; the excerpt does not provide details for assessing the or how the reported rates were calculated.
Open question
What changes would let agents reliably revise their strategies to use solutions they have already discovered?
Source-stated
The authors report that agents discover solutions on Terminal-Bench more often than they exploit them.
agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases
Source-stated
On AppWorld, agents often encounter documentation that explicitly says a command returns the complete solution, but exploit it in very few trials.
agents see documentation stating that a command "returns the complete solution to this task" in over 90% of attempts but exploit this in fewer than 7% of trials
Source-stated
The authors identify available tools in the agent scaffold, test-time compute and training data distribution as factors influencing environmental curiosity.
three main factors influencing environmental curiosity: available tools in the agent scaffold, test-time compute, and training data distribution
Source-stated
The authors report that configurations maximizing environmental curiosity also achieve the best performance on the unmodified benchmarks, while jointly optimized agents still ignore discovered solutions in most trials.
configurations that maximize curiosity also achieve the best performance on the unmodified benchmarks. Yet even jointly optimized agents still ignore discovered solutions in the majority of trials
Source and contributor context
Cohere · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Test-time compute
Extra computation a model spends while answering, for example by reasoning longer, rather than during training.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Editorial provenance
Version 1. Publication ID: 29aee32a-58ad-f612-41e5-766ed0690f6c.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T05:48:50.871Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 5c7820ac3ac29a660128665b9294a6706754d45caf182032fd025bb7df74ac6b. Source text hash: 13fa52536d34d2ef7931441e42df5694126caac59c51763c23afcdd08dfd2201.
Revision and source-check history
Version 1 · 2026-11-09T05:53:23.367Z
Leon Engländer and co-authors argue in Cohere’s paper that agents ignore useful discoveries
Cohere’s paper, published April 19, 2026, by Leon Engländer, Sophia Althammer, Ahmet Üstün, Matthias Gallé and Tom Sherborne, argues that current -based agents often fail to act on useful information they discover in their environments. The authors call the missing capability “environmental curiosity”: recognizing and investigating unexpected but relevant observations. Across Terminal-Bench, SWE-Bench and AppWorld, they exposed agents to complete task solutions. Agents found solutions in many Terminal-Bench runs, but used them much less often; in AppWorld, they often saw documentation explicitly identifying a command’s solution, yet rarely exploited it. The authors identify available tools in the , and training data distribution as factors affecting environmental curiosity. They report that configurations maximizing curiosity also perform best on the unmodified , but even jointly optimized agents ignore discovered solutions in most trials. Their argument is that agents use environments to fetch expected information without reliably revising their strategy when useful, unexpected evidence appears. The authors’ results point to a gap between noticing a solution and adapting behavior to use it; how to close that gap across tasks remains unknown.
Original source
Cohere · Article
Source details
Original source date: Apr 19, 2026 (2026-04-19T00:00:00.000Z).
Date basis: Date stated on the publisher’s page.
Added to Escape Meridian: 2026-11-09T05:53:23.367Z. First observed: 2026-11-09T03:44:12.661Z. Retrieved: 2026-11-09T03:45:36.558Z.
Limits
Text as retrieved on November 9, 2026; originally published April 19, 2026. The supplied evidence is the publisher page’s abstract, not the paper’s full methods or results; the excerpt does not provide details for assessing the or how the reported rates were calculated.
Open question
What changes would let agents reliably revise their strategies to use solutions they have already discovered?
Source-stated
The authors report that agents discover solutions on Terminal-Bench more often than they exploit them.
agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases
Source-stated
On AppWorld, agents often encounter documentation that explicitly says a command returns the complete solution, but exploit it in very few trials.
agents see documentation stating that a command "returns the complete solution to this task" in over 90% of attempts but exploit this in fewer than 7% of trials
Source-stated
The authors identify available tools in the agent scaffold, test-time compute and training data distribution as factors influencing environmental curiosity.
three main factors influencing environmental curiosity: available tools in the agent scaffold, test-time compute, and training data distribution
Source-stated
The authors report that configurations maximizing environmental curiosity also achieve the best performance on the unmodified benchmarks, while jointly optimized agents still ignore discovered solutions in most trials.
configurations that maximize curiosity also achieve the best performance on the unmodified benchmarks. Yet even jointly optimized agents still ignore discovered solutions in the majority of trials
Source and contributor context
Cohere · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Agent harness
An agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Test-time compute
Extra computation a model spends while answering, for example by reasoning longer, rather than during training.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Editorial provenance
Version 1. Publication ID: 29aee32a-58ad-f612-41e5-766ed0690f6c.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: passed by anthropic/claude-opus-5.5 on 2026-11-09T05:48:50.871Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 5c7820ac3ac29a660128665b9294a6706754d45caf182032fd025bb7df74ac6b. Source text hash: 13fa52536d34d2ef7931441e42df5694126caac59c51763c23afcdd08dfd2201.
passed · 2026-11-09T05:48:50.871Z · anthropic/claude-opus-5.5
Publication: 29aee32a-58ad-f612-41e5-766ed0690f6c
attribution — pass: The reading attributes the paper to Cohere and names the five authors listed on the publisher page. Its claims stay framed as the authors' findings. The page is a single paper abstract; the related-works list is a sidebar the reading does not use.
Leon Engländer, Sophia Althammer, Ahmet Üstün, Matthias Gallé, Tom Sherborne
summary — pass: The summary accurately conveys the thesis: agents see injected solutions but rarely use them. It covers the three benchmarks, the three factors, the link between curiosity and performance, and the force of the conclusion. Environmental curiosity is explained in plain words. Scaffold, test-time compute, LLM and benchmark are defined beside the reading. The title and question follow directly from the source's unresolved gap and do not lead the reader.
current agents use the environment to fetch expected information, but not to revise their strategy or maximally exploit useful stimuli
date — pass: The page dateline, Apr 19, 2026, matches evidence.originalPublished, and the reading states the date at day precision. modifiedAt is the same day, so there is no indication of a later edit. This is a publisher page, so no archive disclosure is required. The reading contains no hindsight framing and does not rely on the related-works sidebar.
Apr 19, 2026
support — pass: Each claim matches a verbatim abstract passage: the Terminal-Bench rates, the AppWorld rates, the three factors, and the result on unmodified benchmarks with the jointly optimized caveat. The closing interpretation stays within the source, keeps the unknown open, and adds no invented benefits or costs.
While agents discover these solutions on Terminal-Bench in 79-81% of runs, they interact, or exploit, them in only 37-50% of cases.
media — pass: No media candidates were supplied, and the reading selects none. This is consistent with the evidence.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A language model with many learned numerical settings, used to process and generate language.
Definition referenceAn agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Definition referenceExtra computation a model spends while answering, for example by reasoning longer, rather than during training.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceAn agent harness is the system that lets an AI model act as an agent: it processes inputs, orchestrates tool calls and returns results. An evaluation harness runs evals end to end, including grading outputs.
Definition referenceExtra computation a model spends while answering, for example by reasoning longer, rather than during training.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference