AgentSWE
Coding agents build, edit, or optimize agents across 25 software-engineering commissions. Frozen deliveries run on held-out cases, where execution gates and task-specific model-judged rubrics score behavior. Optimization tasks measure improvement against each source benchmark objective.
Creation and Editing have ten tasks each, while Optimization has five. Each Creation or Editing task has two development cases and six held-out cases. Rubric scores and optimization gap closure use different scales. Some upstream assets remain subject to their original licenses.
- Made by
- Beijing Academy of Artificial Intelligence, Institute of Automation of the Chinese Academy of Sciences, and Renmin University of China
- Released
- October 2026
- Kind
- Machine learning engineering
- Level
- Expert
- Scored
- Graded by a model
- Answers
- Agent task
- Tasks
- All public
- Measures
- Capability
- Language
- English
- Interaction
- Agentic
- Environment
- Terminal
- Refresh
- One fixed set
- Run by
- Anyone
- In views
- Agents
- Strongest model tested
- GPT-5.6 Sol #14 on the capability index
- Last model added
- July 2026
- Scores refreshed
- November 20th, 2026
In the authors’ words
As the production of software passes into the hands of software itself, coding agents are increasingly commissioned to build agents fitted to their users' own workflows. This delegation carries a quiet assumption, that the delivered agent will faithfully serve the requirement it was built for, rather than merely appear to. No existing benchmark tests that assumption. We introduce AgentSWE, a benchmark for agent software engineering spanning three stages of the agent development lifecycle: Creation of a complete agent from a natural-language requirement, Editing of production agent codebases, and Optimization toward a measurable target. Each of its 25 tasks is framed as a commission rather than an exam: the builder iterates against a handful of illustrative cases, while acceptance exercises the frozen submission on held-out cases designed to span the full requirement, grading behavior against requirement-derived rubrics under hard pass/fail gates or, for Optimization, against the source benchmark's own objective. Evaluating a range of coding agents across models and harnesses, we find that none comes close to delivering what a requirement asks for: the gap between what a builder demonstrates and what it delivers is the norm, not the corner case. An accompanying empirical analysis of the failure trajectories identifies recurring failure modes, among them overfitting to the visible cases, hollow runtimes behind code that looks complete, and unsupported claims of completion, each a signal by which coding agents can learn to build the agents they are asked for.
DeepSeek-V4-Pro leads with 19.1 among models released by July 2026
Best score per model from the benchmark's own leaderboard, read on November 20th, 2026. The authors report no overall score, so the headline is Creation, the first stage in their results table. Where a model ran with several settings or harnesses, we keep its best.
The newest model with a score here came out in July 2026, and nobody has tested a newer one. Among the 20 highest-ranked models on the capability index, these have no score yet:
- GPT-6.2 Astra
- Claude Fable 5.2
-
Gemini 4 Carbon
- GPT-6 Astra
- and 15 more
- 1
DeepSeek-V4-Pro 19.1
- 2
DeepSeek-V4-Flash 15.1
- 3 GPT-5.6 Sol 13.7
- 4 GPT-5.5 5.80
- 5
Qwen 3.6 35B-A3B 2.30
Source: Official AgentSWE leaderboard
| # | Model | Score | Setting | Date | Source |
|---|---|---|---|---|---|
| 1 | | 19.1 | Codex | Apr 2026 | Official leaderboard |
| 2 | | 15.1 | Codex | Apr 2026 | Official leaderboard |
| 3 | GPT-5.6 Sol | 13.7 | Codex | Jul 2026 | Official leaderboard |
| 4 | GPT-5.5 | 5.80 | Codex | Apr 2026 | Official leaderboard |
| 5 | | 2.30 | Codex | Apr 2026 | Official leaderboard |
Rankings by part
We read 28 parts from the benchmark’s own leaderboard in 3 groups: creation, editing and optimization. Each ranks models on one part of the test. DeepSeek-V4-Pro leads 10 of them, and 4 other models lead the rest. Arrows show how far a model moved from its overall rank among the models we track.
Authorized vulnerability validation
5 models we track have a score here.
- 1
DeepSeek-V4-Flash ↑1 14.0
- 2
DeepSeek-V4-Pro ↓1 14.0
- 3
Qwen 3.6 35B-A3B ↑2 0.00
- 4 GPT-5.5 0.00
- 5 GPT-5.6 Sol ↓2 0.00
Database analytics
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 14.3
- 2 GPT-5.5 ↑2 9.80
- 3 GPT-5.6 Sol 3.00
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
Desktop GUI automation
5 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 18.0
- 2
DeepSeek-V4-Pro ↓1 17.5
- 3
DeepSeek-V4-Flash ↓1 12.0
- 4 GPT-5.5 3.20
- 5
Qwen 3.6 35B-A3B 0.00
Document-to-editable PPTX
5 models we track have a score here.
- 1
Qwen 3.6 35B-A3B ↑4 21.7
- 2
DeepSeek-V4-Pro ↓1 18.7
- 3
DeepSeek-V4-Flash ↓1 16.8
- 4 GPT-5.5 9.80
- 5 GPT-5.6 Sol ↓2 0.00
Evidence-grounded document QA
5 models we track have a score here.
- 1 GPT-5.5 ↑3 12.2
- 2
DeepSeek-V4-Flash 3.50
- 3 GPT-5.6 Sol 3.30
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Pro ↓4 0.00
Formal theorem proving
5 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 66.7
- 2
DeepSeek-V4-Flash 63.0
- 3
DeepSeek-V4-Pro ↓2 16.5
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5 GPT-5.5 ↓1 0.00
Repository bug repair
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 60.0
- 2 GPT-5.6 Sol ↑1 27.7
- 3
Qwen 3.6 35B-A3B ↑2 0.00
- 4 GPT-5.5 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
Schema-guided web extraction
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 6.70
- 2
Qwen 3.6 35B-A3B ↑3 0.00
- 3 GPT-5.5 ↑1 0.00
- 4
DeepSeek-V4-Flash ↓2 0.00
- 5 GPT-5.6 Sol ↓2 0.00
Scientific PDF translation
5 models we track have a score here.
- 1
DeepSeek-V4-Flash ↑1 26.0
- 2
DeepSeek-V4-Pro ↓1 25.3
- 3 GPT-5.5 ↑1 21.8
- 4 GPT-5.6 Sol ↓1 18.3
- 5
Qwen 3.6 35B-A3B 0.00
Web research report
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 18.0
- 2
DeepSeek-V4-Flash 15.3
- 3
Qwen 3.6 35B-A3B ↑2 1.70
- 4 GPT-5.5 0.80
- 5 GPT-5.6 Sol ↓2 0.00
Creation
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 19.1
- 2
DeepSeek-V4-Flash 15.1
- 3 GPT-5.6 Sol 13.7
- 4 GPT-5.5 5.80
- 5
Qwen 3.6 35B-A3B 2.30
Claude policy provenance
5 models we track have a score here.
- 1 GPT-5.5 ↑3 43.7
- 2
DeepSeek-V4-Pro ↓1 28.3
- 3 GPT-5.6 Sol 24.0
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
Aider worktree transaction
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 47.3
- 2 GPT-5.6 Sol ↑1 21.5
- 3 GPT-5.5 ↑1 20.0
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
OpenHands effect recovery
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 24.3
- 2 GPT-5.5 ↑2 20.5
- 3 GPT-5.6 Sol 4.00
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
OpenClaw channel handoff
5 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 38.5
- 2
DeepSeek-V4-Pro ↓1 19.5
- 3 GPT-5.5 ↑1 12.3
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
Codex execution residual
5 models we track have a score here.
- 1 GPT-5.5 ↑3 70.0
- 2
DeepSeek-V4-Pro ↓1 35.0
- 3 GPT-5.6 Sol 11.7
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
AI-Scientist reproducibility gate
5 models we track have a score here.
- 1 GPT-5.5 ↑3 67.3
- 2
DeepSeek-V4-Pro ↓1 61.7
- 3
DeepSeek-V4-Flash ↓1 54.3
- 4 GPT-5.6 Sol ↓1 48.5
- 5
Qwen 3.6 35B-A3B 0.00
DeepCode claim traceability
5 models we track have a score here.
- 1
DeepSeek-V4-Flash ↑1 47.0
- 2 GPT-5.5 ↑2 41.3
- 3
DeepSeek-V4-Pro ↓2 32.7
- 4 GPT-5.6 Sol ↓1 26.5
- 5
Qwen 3.6 35B-A3B 0.00
DeepTutor adaptive remediation
5 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 54.2
- 2
DeepSeek-V4-Pro ↓1 31.0
- 3
DeepSeek-V4-Flash ↓1 25.7
- 4 GPT-5.5 14.3
- 5
Qwen 3.6 35B-A3B 0.00
Dyad acceptance-driven workflow
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 22.5
- 2 GPT-5.5 ↑2 18.0
- 3 GPT-5.6 Sol 9.80
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5
DeepSeek-V4-Flash ↓3 0.00
OpenWiki change-impact documentation
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 14.0
- 2
DeepSeek-V4-Flash 12.8
- 3 GPT-5.5 ↑1 8.50
- 4
Qwen 3.6 35B-A3B ↑1 0.00
- 5 GPT-5.6 Sol ↓2 0.00
Editing
5 models we track have a score here.
- 1 GPT-5.5 ↑3 31.6
- 2
DeepSeek-V4-Pro ↓1 31.6
- 3 GPT-5.6 Sol 23.9
- 4
DeepSeek-V4-Flash ↓2 14.0
- 5
Qwen 3.6 35B-A3B 0.00
BrowseComp
5 models we track have a score here.
- 1
DeepSeek-V4-Flash ↑1 34.1%
- 2 GPT-5.6 Sol ↑1 32.3%
- 3 GPT-5.5 ↑1 18.4%
- 4
Qwen 3.6 35B-A3B ↑1 8.50%
- 5
DeepSeek-V4-Pro ↓4 -1.00%
τ³
5 models we track have a score here.
- 1
DeepSeek-V4-Pro 15.8%
- 2 GPT-5.6 Sol ↑1 4.30%
- 3 GPT-5.5 ↑1 0%
- 4
Qwen 3.6 35B-A3B ↑1 -23.80%
- 5
DeepSeek-V4-Flash ↓3 -50.00%
Terminal-Bench
5 models we track have a score here.
- 1
Qwen 3.6 35B-A3B ↑4 7.70%
- 2 GPT-5.5 ↑2 7.70%
- 3
DeepSeek-V4-Flash ↓1 2.80%
- 4
DeepSeek-V4-Pro ↓3 0%
- 5 GPT-5.6 Sol ↓2 -10.00%
PinchBench
5 models we track have a score here.
- 1
Qwen 3.6 35B-A3B ↑4 12.1%
- 2 GPT-5.5 ↑2 -11.10%
- 3 GPT-5.6 Sol -14.90%
- 4
DeepSeek-V4-Flash ↓2 -34.40%
- 5
DeepSeek-V4-Pro ↓4 -38.70%
OSWorld
5 models we track have a score here.
- 1
DeepSeek-V4-Flash ↑1 12%
- 2 GPT-5.6 Sol ↑1 12%
- 3
DeepSeek-V4-Pro ↓2 4.00%
- 4
Qwen 3.6 35B-A3B ↑1 0%
- 5 GPT-5.5 ↓1 -9.10%
Optimization
5 models we track have a score here.
- 1 GPT-5.6 Sol ↑2 4.70%
- 2 GPT-5.5 ↑2 1.20%
- 3
Qwen 3.6 35B-A3B ↑2 0.90%
- 4
DeepSeek-V4-Pro ↓3 -4.00%
- 5
DeepSeek-V4-Flash ↓3 -7.10%
Number one over time
Lab claims count from the day the card came out, independent runs from the model’s release.
Source: Official AgentSWE leaderboard
Scores over time
All the scores we have, one dot each, so a model tested at several settings appears more than once.
Source: Official AgentSWE leaderboard
No lab model card reports it yet
We read the headline results of 540 launch posts from 57 labs, and none of them list this benchmark.
More in AI research
- PaperBench In use Agent replicates ICML 2024 papers from scratch. LLM judge grades against author-written hierarchical rubrics.
Qwen 3.8 Max 93%
- MLS-Bench In use Agent must invent one algorithmic change in an ML research scaffold whose gain transfers across settings. Scored relative to human baselines. Claude Fable 5.1 50.3%
- PostTrainBench 1.2 In use CLI agent gets a base model, one H100 and a time budget to post-train it for a target benchmark. Scored on normalized gain. Claude Fable 5.1 44.6%
- MLE-bench In use Kaggle-style ML engineering competitions. Agent trains models and is scored against competition leaderboards. GPT-6.2 Astra 96.1%
- PMPP-Hard In use Agents write GPU kernels for 69 tasks that must be correct, and on 48 of them also run within 1.25 times the reference time. The score is the area under the solved-tasks curve over a 10-hour budget, out of 690 task-hours. Kimi K3 0.71
- KernelBench In use Replace PyTorch operators with faster custom GPU kernels. Scored on correctness and speedup.
MiniMax-M3 28.8%
Data audit
All 9 checks pass. Every benchmark
Catalog record 2 pass
- Pass Task access All tasks are public.
- Pass Confidence Checked against a primary source.
Scores 3 pass 2 notes
- Pass Independent runs Scores from the official leaderboard (5 entries).
- Pass Models scored 5 models we track.
- Pass Newest score The newest is from July 2026, for GPT-5.6 Sol.
- Note Frontier models None of the top 10 has a score. The strongest tested is GPT-5.6 Sol (#14).
- Note Headroom Scored as a plain number, with no fixed ceiling.
Leaderboard import 3 pass
- Pass Importer We read the board on vectorspacelab.github.io: 5 rows for models we track at the last read.
- Pass Last read November 20th, 2026.
- Pass Read worked The last read worked.
Consistency 1 pass 1 note
- Note Lab claims No model has both a lab number and an independent run.
- Pass Known issues No published audit found problems.