<!-- mobian-agent-page publisher="dailydev" canonical="https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe" -->

---
title: SWE-CoWork: a coding agent benchmark built around team...
description: SWE-CoWork tests whether coding agents can handle team-style handoffs, not just isolated patches. Agents must gather requirements from tickets and teammates,...
canonical: https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe
twitter:card: summary_large_image
twitter:site: @dailydotdev
og:type: website
og:site_name: daily.dev
og:title: SWE-CoWork: a coding agent benchmark built around team handoffs | daily.dev
og:description: SWE-CoWork tests whether coding agents can handle team-style handoffs, not just isolated patches. Agents must gather requirements from tickets and teammates,...
og:url: https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe
og:image: https://api.daily.dev/og/posts/dWpkuTuxE.png
og:image:alt: SWE-CoWork: a coding agent benchmark built around team handoffs
og:image:width: 1200
og:image:height: 630
og:locale: en
---

> ## Documentation Index
> Fetch the complete documentation index at: https://daily.dev/llms.txt
> Use this file to discover all available pages before exploring further.

# SWE-CoWork: a coding agent benchmark built around team handoffs

**[Collections](https://daily.dev/sources/collections)** · 2 min read · 0 upvotes · 0 comments

## Summary

SWE-CoWork tests whether coding agents can handle team-style handoffs, not just isolated patches. Agents must gather requirements from tickets and teammates, write code, verify it, and report results across 101 real repositories. The best model fully finished only 8 of 101 tasks. For Claude Opus 5.5, more than half its failed checks happened after it already had the requirement, suggesting the weak point is carrying information through to working code rather than gathering it. @svpino notes Opus 5.5 is at the top of the leaderboard, with GPT-6 Astra a close second.

## Content

TokenWave AI has released SWE-CoWork, a benchmark that asks a different question from most coding evaluations. Instead of handing an agent a complete written task, it drops the agent into a simulated workplace with tickets, docs, and teammates it can ask. About 82% of the requirements sit with those teammates, so the agent only learns what to build by asking the right person.

The agent then has to deliver a patch that passes hidden tests across 101 real repos. Each failed check is linked to a specific requirement, which makes it possible to see where things went wrong.

As @svpino describes it, the benchmark measures whether agents can be dependable digital engineers on long-horizon tasks. To pass, an agent has to gather requirements, coordinate with colleagues, write the code, verify its work, and report its results.

## Results

The numbers are humbling. The best models pass about 63% of individual checks but fully finish fewer than 8% of the 101 tasks. The top model completed 8 of them.

On the leaderboard, Claude Opus 5.5 is at the top, with GPT-6 Astra a close second.

Two details stand out:

- Giving GPT-6 Astra the full spec up front lifted its passed checks by 16 points, yet it still fully finished only 8 of 101 tasks.
- For Claude Opus 5.5, more than half of the failed checks came after the agent already had the requirement. So the problem isn't only knowing who to ask. Even with the information in hand, the agent often fails to carry it through.

## Why it matters

@omarsar0 says he is watching handoff capabilities closely, and notes they are still hard to build into agents. This benchmark measures exactly that. It reads a bit like a new hire's first week: lots of context scattered across people and documents, and a lot of ways to miss something.

I find the second finding more interesting than the headline score. Asking the right question is one skill. Remembering and applying the answer across a whole patch is another, and the data suggests that is where agents still slip.

## Questions this post answers

### What is the SWE-CoWork benchmark for coding agents?

SWE-CoWork is a benchmark that tests whether coding agents can work like dependable engineers on long tasks instead of just writing an isolated patch. Agents are placed in a simulated team with tickets, docs, and teammates to question, and must gather requirements, coordinate, write code, verify it, and report results. The final patch must pass hidden tests across 101 real repositories. It comes from TokenWaveai.

_Tracking how agents handle real handoffs, not toy problems, is easier with daily.dev's view of agent benchmarks._

### How well does Claude Opus 5.5 perform on the SWE-CoWork agent benchmark?

Claude Opus 5.5 is the best-performing model on SWE-CoWork but fully completes only 8 of 101 tasks. More than half of its failed checks occurred after the agent already had the requirement, showing the bottleneck is carrying information through to working code rather than gathering it. Opus 5.5 currently tops the leaderboard, with GPT-6 Astra a close second.

_Developers picking an agent for real handoff work can follow benchmark results like this on daily.dev._

## Community take

How the wider developer community reacted, aggregated from 3 discussions and 69 comments across x (as of 2026-11-18).

**TL;DR:** Commenters largely agree the benchmark's real finding is that requirements-gathering isn't the main bottleneck for coding agents; the bigger problem is losing track of or failing to act on requirements they already have, especially during verification and handoff.

**Sentiment:** 10% positive · 55% mixed · 35% skeptical

**The case for**

- The benchmark's design of scoring requirement-gathering, coordination, verification, and reporting as separate steps is seen as a meaningful and realistic test of teamwork, not just coding.
- Several see the oversight/tracking failure mode as mirroring real human engineering handoffs, which makes the benchmark feel credible.
- Some suggest practical fixes like writing confirmed requirements into a file or ticket and rechecking them before handoff, implying the benchmark usefully exposes where to intervene.

**The pushback**

- Many argue that giving agents the full spec upfront raised test-pass rates but didn't move full-task completion, suggesting the benchmark's framing around asking teammates may overstate that bottleneck.
- Commenters distrust self-reported 'done' status, noting agents often claim tests passed without actually running them or quietly skip checks.
- Some question how the 'gather requirements' and 'coordinate with colleagues' steps are actually scored, since another model plays the teammates and controls what's shared.
- A few doubt whether the verification/reporting gap is really about memory versus genuine inability to check work.

**By community**

- x (mixed): Replies converge on the idea that keeping and acting on a known requirement, not gathering it, is the real failure point, while still debating whether that's a memory, oversight, or verification problem.

**Hottest debate:** Whether the core bottleneck is agents failing to ask the right questions (handoff/requirements gathering) or failing to retain and act on requirements they already have (tracking/verification).

**Open questions**

- Does the benchmark show exactly where a known requirement gets dropped — during planning or during the actual code edit?
- How is 'gather requirements' or 'coordinate with colleagues' scored when another model plays the teammates controlling what's shared?
- How many failed checks trace to a question never asked versus an answer the agent received and then lost?
- Were other prominent models like Grok or Muse evaluated, or excluded from the leaderboard?

**Highlights**

> @omarsar0 The gap is the story: GPT-6 Astra and Opus 5.5 pass ~63% of new-behavior tests but fully finish just 8 and 7 of 101 tasks. And in Opus's 979 traced failures, 56% came after it already had the requirement. Better question-asking fixes less than half the handoff problem.
> — [AIRoboticsInt on x](https://x.com/AIRoboticsInt/status/2122723771576680524)

> @omarsar0 from the paper, asking isn't the main gap: 56% of failed linked tests were requirements already disclosed, and 153 of 450 deferred questions were never followed up (median 7.1h budget left). a requirements file the agent re-reads after each compaction would help more
> — [victor\_explore on x](https://x.com/victor_explore/status/2122727777938092264)

> @rohanpaul_ai A new hire who shipped 8 of 101 tickets would still get a second week, to be fair. What gets me is the full spec not moving that 8 at all, so asking teammates isn't really the bottleneck.
> — [AIQuanting on x · 1 points](https://x.com/AIQuanting/status/2122733892541276271)

> @svpino Step 4 is where agents write "All tests pass ✅" without running a single one. Cheap guard: have the harness check the exit code, not the summary. pytest -q; echo $? 0 or it didn't happen.
> — [termuxSolutions on x](https://x.com/termuxSolutions/status/2122722721075523597)

> @rohanpaul_ai 82% of requirements held by teammates: that's a plant project too. Of the failed checks, how many trace to a question never asked, and how many to an answer the agent got and then lost?
> — [bobbyships on x](https://x.com/bobbyships/status/2122737964241535215)

**Source threads**

- [x](https://x.com/omarsar0/status/2122721966574408112) · 0 points · 25 comments
- [x](https://x.com/svpino/status/2122718723664720009) · 0 points · 33 comments
- [x](https://x.com/rohanpaul_ai/status/2122730768317476944) · 0 points · 11 comments

## Similar posts on daily.dev

- [AI’s best coding agent fails 60% of the time — and the data backs it up](https://daily.dev/posts/ai-s-best-coding-agent-fails-60-of-the-time-and-the-data-backs-it-up-ziebak7ti) · The New Stack · 2 upvotes · 0 comments
- [13 Models and 4 Agents on SWE Tasks: Go, Java, Python, Rust, TS](https://daily.dev/posts/13-models-and-4-agents-on-swe-tasks-go-java-python-rust-ts-ithomlqxu) · Hacker News · 2 upvotes · 0 comments

---

Tags: [#ai](https://daily.dev/tags/ai), [#ai-agents](https://daily.dev/tags/ai-agents), [#claude](https://daily.dev/tags/claude)

[View this post on daily.dev](https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe)

```json
{"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]}
{"@context":"https://schema.org","@type":"TechArticle","headline":"SWE-CoWork: a coding agent benchmark built around team handoffs","url":"https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe"},"datePublished":"2026-11-17T16:02:42.657Z","dateModified":"2026-11-18T05:57:00.536Z","description":"SWE-CoWork tests whether coding agents can handle team-style handoffs, not just isolated patches. Agents must gather requirements from tickets and teammates,...","image":"https://pbs.twimg.com/media/HXVrlZ3WcAIrMGc.jpg","thumbnailUrl":"https://pbs.twimg.com/media/HXVrlZ3WcAIrMGc.jpg","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai,ai-agents,claude","timeRequired":"PT2M"}
{"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"SWE-CoWork: a coding agent benchmark built around team handoffs"}]}
{"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/swe-cowork-a-coding-agent-benchmark-built-around-team-handoffs-dwpkutuxe#faq","mainEntity":[{"@type":"Question","name":"What is the SWE-CoWork benchmark for coding agents?","acceptedAnswer":{"@type":"Answer","text":"SWE-CoWork is a benchmark that tests whether coding agents can work like dependable engineers on long tasks instead of just writing an isolated patch. Agents are placed in a simulated team with tickets, docs, and teammates to question, and must gather requirements, coordinate, write code, verify it, and report results. The final patch must pass hidden tests across 101 real repositories. It comes from TokenWaveai. Tracking how agents handle real handoffs, not toy problems, is easier with daily.dev's view of agent benchmarks."}},{"@type":"Question","name":"How well does Claude Opus 5.5 perform on the SWE-CoWork agent benchmark?","acceptedAnswer":{"@type":"Answer","text":"Claude Opus 5.5 is the best-performing model on SWE-CoWork but fully completes only 8 of 101 tasks. More than half of its failed checks occurred after the agent already had the requirement, showing the bottleneck is carrying information through to working code rather than gathering it. Opus 5.5 currently tops the leaderboard, with GPT-6 Astra a close second. Developers picking an agent for real handoff work can follow benchmark results like this on daily.dev."}}]}
```

