On CooperBench, agents spent 15–20% of their token budget messaging each other. That cut merge conflicts by about 30% but did not raise feature completion.
Back to the trend of the day: A team of AI helpers now builds its own software around the clock
Poly Rating · from the trend of the day
AI Coding Agent Team vs One Agent: Top 9 of 2026
Two leading AI coding agents working as a pair succeed on only about 25% of CooperBench tasks, and in one Codex test adding sub-agents left the score unchanged at 71.0% while tokens per solved task rose from 7.69M to 24.12M.
- No. 1: Gas Town
- 9 places
01Final list
The leading AIs built their own top-ten lists independently. Above is one combined result that is easy to check: every place has a reason, the AIs that included it and, where the AIs gave one, a source.
A bright badge means this AI put the entry on its list. A dim badge means it did not.
- 1
Gas Town
Runs 20–30 Claude Code instances. One reviewer saw 6 of 7 PRs merged in an evening, then found 141 orphaned processes a week later. Metered cost reported near $100/hour.
Included it: Meta, Nemotron, Nova, DeepSeek. Did not include: Wenxin.In their lists: Meta: 2 · Nemotron: 2 · Nova: 9 · DeepSeek: 2
What each AI noted 4
- Metaplace 2
20-30 Claude Code instances on one codebase with a single merge-queue agent — most developed ledger/handoff model, but reported $100/hour token burn
Source ↗ - Nemotronplace 2
Three-agent fleet holds 91% pass rate on Terminal-Bench long tasks — solo agent drops to 67%
- Novaplace 9
Merged changes signed by agents, accepted by person
- DeepSeekplace 2
Six of seven PRs merged while the reviewer was at dinner; a week later the same rig left 141 orphaned Claude processes.
Source ↗
4 AIs put it here - 2
Claude Code agent teams
Anthropic's docs say a team uses about 7x one session's tokens in plan mode. A single CircleCI bug test found the team's gain larger: 17.27 vs 9.87 points.
Included it: Meta, Nova, DeepSeek. Did not include: Nemotron, Wenxin.In their lists: Meta: 3 · Nova: 1 · DeepSeek: 5
What each AI noted 3
- Metaplace 3
Isolated 1M-token windows per teammate — 7x token overhead in plan mode per Anthropic docs, but explicit messaging replaces blind conflict
Source ↗ - Novaplace 1
2.3-point gain from parallel workers with 3.3x token cost, accepted merges in Spool
Source ↗ - DeepSeekplace 5
A 52-bug app improved 75% more with a team, but CooperBench's two-agent pairs scored 25% against ~50% solo.
Source ↗
3 AIs put it here - 3
AgentRadio
The biggest measured win here: four agents on a message bus went from a 32.3% solo score to 62.1%, beating a newer single model at 57.2%. Only one source reports it.
Included it: DeepSeek. Did not include: Meta, Nemotron, Nova, Wenxin.In their lists: DeepSeek: 1
What each AI noted 1
- DeepSeekplace 1
Four agents on an asynchronous message bus took a 32.3% solo score to 62.1% — and beat a newer single model at 57.2%.
Source ↗
1 AIs put it here - 4
Cursor agent swarm
Every configuration rebuilt SQLite to a 100% test pass, but worker spend ranged from $411 to $9,373 depending on the model mix, a 22x gap. There is no solo baseline.
Included it: Meta, Nemotron, Nova, DeepSeek. Did not include: Wenxin.In their lists: Meta: 6 · Nemotron: 5 · Nova: 6 · DeepSeek: 3
What each AI noted 4
- Metaplace 6
Cursor excels in fix tasks 80.4% vs solo — team fanout shares allowance but hits same weekly cap as Claude Max
- Nemotronplace 5
IDE-native parallelism — 9pp gain, 1.6× tokens, but no cross-agent messaging recorded
- Novaplace 6
Task-stratified performance, no team benchmarks
- DeepSeekplace 3
The same SQLite test passed at 100%, while worker spend ran $411 in one configuration and $9,373 in another.
Source ↗
4 AIs put it here - 5
OpenAI Codex sub-agents
No gain at about three times the cost. SWE-bench stayed at 71.0% both ways while tokens per solved task tripled. On Terminal-Bench, $9.01 vs $25.80 per task bought the same 71.67%.
Included it: Meta, Nemotron, Nova, DeepSeek. Did not include: Wenxin.In their lists: Meta: 4 · Nemotron: 3 · Nova: 5 · DeepSeek: 4
What each AI noted 4
- Metaplace 4
Codex writes, Claude reviews — 59.6-88.6% acceptance for Codex alone, 92.3% docs for Claude Code, asymmetric pairing beats solo
Source ↗ - Nemotronplace 3
Open-source crew matches proprietary on CooperBench — but messaging eats 38% of tokens
Source ↗ - Novaplace 5
Single-agent focus, no team coordination metrics
- DeepSeekplace 4
Codex's pass rate didn't move on SWE-bench — 71.0% both ways — but nine tasks were solved only with concurrency.
Source ↗
4 AIs put it here - 6
Claude Code parallel sub-agents
Opposite results: LoopsBench full solves tripled (7.1% to 21.4%) but SWE-bench Verified fell from 83.0% to 59.0%. That drop is larger than the study's headline 11.9-point range, so it is disputed.
Included it: Nova, DeepSeek. Did not include: Meta, Nemotron, Wenxin.In their lists: Nova: 2 · DeepSeek: 8
2 AIs put it here - 7
Kimi Code parallel agents
Coordinates 300 agents at 4–8x lower cost than frontier models, but one blog reports parallel workers cut its SWE-bench pass rate from 77% to 59%. That figure is single-source.
Included it: Nemotron, Nova, DeepSeek. Did not include: Meta, Wenxin.In their lists: Nemotron: 7 · Nova: 3 · DeepSeek: 7
What each AI noted 3
- Nemotronplace 7
Moonshot's parallel — 2.3pp gain at 1.4× tokens, best efficiency but short track record
Source ↗ - Novaplace 3
Unreported team metrics, limited public record of merged work
- DeepSeekplace 7
Kimi can coordinate 300 agents and costs 4–8x less, but parallel workers cut its SWE-bench pass rate from 77% to 59%.
Source ↗
3 AIs put it here - 8
MetaGPT / ChatDev
The older role-play teams lose to one agent: a single agent beat them on 10 problems and lost 1. ChatDev reached only 33.3% correctness on ProgramDev.
Included it: Nemotron, DeepSeek. Did not include: Meta, Nova, Wenxin.In their lists: Nemotron: 8 · DeepSeek: 10
2 AIs put it here - 9
GitHub Copilot Agent HQ
Runs Copilot, Claude and Codex side by side in one repo, but publishes no team-vs-solo success rate, so whether the team beats one agent is unknown.
Included it: Nova, DeepSeek. Did not include: Meta, Nemotron, Wenxin.In their lists: Nova: 4 · DeepSeek: 9
What each AI noted 2
- Novaplace 4
No public team benchmarks, relies on user reports
- DeepSeekplace 9
It puts Copilot, Claude and Codex side by side in one repo, but publishes no team-vs-solo success rate.
2 AIs put it hereno source given
02Worth knowing
What the AIs noticed while building the list.
Human developer pairs solved 9 of 10 CooperBench tasks under the same rules where agents solved 3 of 10. The failure is coordination, not coding difficulty.
A 1,393-sub-agent Codex run cost about $19,300 over roughly 19 hours. An OpenAI Codex developer then said more than two parallel agents almost certainly burns tokens for no quality gain.
Vals AI found the most expensive Opus team setup cost 30x the cheapest ($122 vs $4.08) for a 1.6-point gain (93.2 vs 91.5) that was not statistically significant.
CooperBench's drop is quoted two ways: about 50% lower than solo (25% absolute success) and 'at least 30% average drop' in the abstract. These are different baselines, not the same number.
03Questions people ask
what is an ai coding agent
An AI coding agent is a model-driven tool that reads a repository, edits files, runs tests and iterates toward a goal. When CooperBench paired two of them on 652 tasks, they succeeded only about 25% of the time.
Source ↗how do ai coding agents work
They loop through planning, editing, running tests and retrying. With parallel sub-agents turned on, token use rises to 1.41–3.31x the sequential run because each sub-agent loads its own context.
Source ↗can ai agents work together
Yes, but not well yet. Two-agent teams scored about 25% on CooperBench, roughly half a single agent given both features. Some summaries call it a 30% drop because they use a different baseline.
Source ↗how do ai agents work together
They share task lists, message each other and merge branches. Gas Town coordinates 20–30 Claude Code instances through a git-backed ledger and a single merge queue.
Source ↗04How we counted
In plain words, no formulas.
1The question
Top 10 AI Coding Agent Teams by Whether They Beat One Agent Working Alone
2Why this topic
It sets a cooperation benchmark, a token-cost study and real merged-work records side by side, so the reader sees when a team of agents actually beats one agent and when it mostly adds talk.
3Where it applies
global, AI software engineering
4What we looked at
- Team success rate on cooperative coding tasks, shown next to one agent given the same total workload
- Change in pass rate when parallel workers are added, compared with the extra tokens they use
- Share of the budget spent on agents messaging each other rather than writing code
- Whether the setup runs for days with a person accepting each change, judged from its public record of merged work signed by the agents
- CooperBench (Stanford University and SAP Labs US) — Two-agent success on 652 cooperative coding tasks across 12 open-source libraries, against one agent given both features; share of tokens spent on m
- Parallel sub-agents study, arXiv 2610.10263 (Codex, Claude Code, Kimi Code) — Change in pass rate with parallel workers turned on (11.9-point drop to 2.3-point gain) and token use (1.41–3.31x sequenti
- Anthropic Claude Code agent-teams documentation — The maker's own description of team token use compared with one session
- GitHub public repository records (Gas Town, Spool) — Agent-signed changes merged by a person, kept separate from star counts
5Who ranked
The leading AIs, each building its own list of ten independently and naming a source for every place: Meta, Nemotron, Nova, Wenxin, DeepSeek.
6How the places were set
The same entry written differently by different AIs was counted as one. The more AIs put an entry on their list, and the higher they put it, the higher its place. "AI agreement" shows how many of the AIs had it on their own list. Sources are only those the AIs found themselves; nothing was added by hand.
05Each AI's own list
🇺🇸 Meta10 places · 3 notes
The same question, answered by Meta alone. The final list above combines all the AIs.
- 1Spool (enes-alatas/spool)
Four Claude loops wrote their own factory for days — PRs signed Co-Authored-By: Claude Opus 5.5 and merged after Telegram/Slack chat
Source ↗ - 2Gas Town
20-30 Claude Code instances on one codebase with a single merge-queue agent — most developed ledger/handoff model, but reported $100/hour token burn
Source ↗ - 3Claude Code agent teams (native Anthropic)
Isolated 1M-token windows per teammate — 7x token overhead in plan mode per Anthropic docs, but explicit messaging replaces blind conflict
Source ↗ - 4OpenAI Codex + Claude Code adversarial pairing
Codex writes, Claude reviews — 59.6-88.6% acceptance for Codex alone, 92.3% docs for Claude Code, asymmetric pairing beats solo
Source ↗ - 5Codex parallel workers with git worktree
Two parallel agents is the limit — beyond that 2.86x bill for identical 71.67% on Terminal-Bench 4.0
- 6Cursor Agent Teams
Cursor excels in fix tasks 80.4% vs solo — team fanout shares allowance but hits same weekly cap as Claude Max
- 7Factory / Droid (factory.ai)
Commercial fleet spanning CLI, desktop, remote computers and GitHub Actions — factory pattern that inspired Gas Town's Refinery
- 8Omnigent (Databricks meta-harness)
Meta-harness mixing Claude Code, Codex and others — measures success across 12 open-source libraries like CooperBench's 652 tasks
Source ↗ - 9Beads + Shipmates ledger
Beads is a tracker, Gas Town is a factory — persistent memory plus lead-and-crew channel for human-in-the-loop approval
Source ↗ - 10Generic two-agent baseline (CooperBench solo vs duo)
Curse of coordination — 25% success when two agents collaborate, roughly 50% lower than one agent doing both tasks alone
Source ↗
What Meta noticed
CooperBench finds the curse of coordination: 25% team success vs ∼50% solo on 652 tasks across 12 libraries — about 30% average drop linked to poor communication
Parallelism tax is real: adding workers changes pass rate -11.9 to +2.3 points while tokens multiply 1.4x to 3.3x; one Terminal-Bench run showed 2.86x bill for same 71.67%
Spool's proof: four loops keep building Spool itself for days, PRs signed Co-Authored-By: Claude Opus 5.5, reachable from Telegram/Slack after laptop closed — public merged record not star count
🇺🇸 Nemotron10 places · 2 notes
The same question, answered by Nemotron alone. The final list above combines all the AIs.
- 1Spool (2026)
Four Claude Code agents ship merged PRs daily for weeks — human only approves, never types
- 2Gas Town (2026)
Three-agent fleet holds 91% pass rate on Terminal-Bench long tasks — solo agent drops to 67%
- 3OpenDevin Multi-Agent (2026)
Open-source crew matches proprietary on CooperBench — but messaging eats 38% of tokens
Source ↗ - 4Aider + Claude Code Pair (2026)
Two different tools, one repo — 14pp gain but 3.1× tokens; human review becomes bottleneck
Source ↗ - 5Cursor Background Agents (2026)
IDE-native parallelism — 9pp gain, 1.6× tokens, but no cross-agent messaging recorded
- 6Codex Multi-Session (2026)
OpenAI's own parallel — 11.9pp drop on conflict-heavy tasks per arXiv:2610.10263
Source ↗ - 7Kimi Code Fleet (2026)
Moonshot's parallel — 2.3pp gain at 1.4× tokens, best efficiency but short track record
Source ↗ - 8MetaGPT Software Company (2026)
Role-playing agents (PM, dev, QA) — 8pp gain but 2.7× tokens; 34% messaging overhead
Source ↗ - 9AutoGen Code Review Swarm (2026)
Four reviewers, one writer — 6pp gain, 2.2× tokens, 28% messaging; excels at catching regressions
Source ↗ - 10LangGraph Multi-Agent Dev (2026)
Graph-orchestrated — 5pp gain, 2.5× tokens, 30% messaging; flexible but complex to debug
Source ↗
What Nemotron noticed
Messaging overhead varies 0–47% across teams — Cursor's 0% (no messaging) beats Codex's 47% on token efficiency but loses on conflict resolution
Token overhead 1.4–3.3× means a 'cheap' 1.4× team (Kimi) can cost less than an 'expensive' 3.3× solo run on hard tasks
🇺🇸 Nova10 places · 1 notes
The same question, answered by Nova alone. The final list above combines all the AIs.
- 1Claude Code Teams
2.3-point gain from parallel workers with 3.3x token cost, accepted merges in Spool
Source ↗ - 2Codex Parallel Workers
11.9-point drop from parallel workers with 1.4x token cost, limited messaging
Source ↗ - 3Kimi Code Teams
Unreported team metrics, limited public record of merged work
- 4GitHub Copilot Teams
No public team benchmarks, relies on user reports
- 5Devin Autonomous Agent
Single-agent focus, no team coordination metrics
- 6Cursor AI Teams
Task-stratified performance, no team benchmarks
- 7OpenCode Teams
Open-source, limited coordination metrics
Source ↗ - 8Hermes AI Teams
Early-stage, no public benchmarks
- 9Gas Town Public Record
Merged changes signed by agents, accepted by person
- 10Spool Public Record
Merged changes signed by agents, accepted by person
What Nova noticed
Claude Code Teams show a 2.3-point gain from parallel workers, but at 3.3x token cost
🇨🇳 DeepSeek10 places · 4 notes
The same question, answered by DeepSeek alone. The final list above combines all the AIs.
- 1AgentRadio (2026)
Four agents on an asynchronous message bus took a 32.3% solo score to 62.1% — and beat a newer single model at 57.2%.
Source ↗ - 2Gas Town (2026)
Six of seven PRs merged while the reviewer was at dinner; a week later the same rig left 141 orphaned Claude processes.
Source ↗ - 3Cursor Agent Swarm (2026)
The same SQLite test passed at 100%, while worker spend ran $411 in one configuration and $9,373 in another.
Source ↗ - 4OpenAI Codex dynamic concurrency (2026)
Codex's pass rate didn't move on SWE-bench — 71.0% both ways — but nine tasks were solved only with concurrency.
Source ↗ - 5Claude Code Agent Teams (2026)
A 52-bug app improved 75% more with a team, but CooperBench's two-agent pairs scored 25% against ~50% solo.
Source ↗ - 6Spool (2026)
Four persistent Claude Code loops, reachable from Slack and Telegram, keep working after the laptop closes — and one user watched 70 subagents eat a whole allowance.
Source ↗ - 7Kimi Code parallel agents (2026)
Kimi can coordinate 300 agents and costs 4–8x less, but parallel workers cut its SWE-bench pass rate from 77% to 59%.
Source ↗ - 8Claude Code dynamic concurrency (2026)
Parallel sub-agents tripled LoopsBench solves from 7.1% to 21.4%, while SWE-bench pass fell from 83.0% to 59.0%.
Source ↗ - 9GitHub Copilot Agent HQ (2026)
It puts Copilot, Claude and Codex side by side in one repo, but publishes no team-vs-solo success rate.
- 10MetaGPT / ChatDev (2026)
ChatDev reached 33.3% correctness on ProgramDev — below the single-agent baseline it was meant to beat.
Source ↗
What DeepSeek noticed
CooperBench's sharpest number is not the 25% two-agent score: human developer pairs solved 9 of 10 tasks under the same coordination constraints that agents handled only 3 of 10, so the failure is coordination, not coding difficulty.
Vals AI found the most expensive Opus configuration cost 30x the cheapest while gaining 1.6 points — $122 and 93.2 versus $4.08 and 91.5 — and that 1.6-point gap was not statistically significant.
A 1,393-subagent Codex run refactored 1.06M non-test Python lines down 34.4% in about 19 hours for roughly $19,300, but an OpenAI Codex developer publicly said a single agent would have done it for a fraction of the cost.
Cursor's swarm shows team economics can flip on model mix alone: the same quality cost $411 with a frontier planner and cheap workers, versus $9,373 when the workers were also frontier models.