Poly Rating · from the trend of the day

AI Coding Agent Team vs One Agent: Top 9 of 2026 · 9 places

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.

Order: from most to least AI agreement.

A bright badge means this AI put the entry on its list. A dim badge means it did not.

  1. 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.

    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
    Source ↗
  2. 2

    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.

    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
    Source ↗
  3. 3

    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.

    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
    Source ↗
  4. 4

    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.

    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
    Source ↗
  5. 5

    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.

    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
    Source ↗
  6. 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.

    In their lists: Nova: 2 · DeepSeek: 8

    What each AI noted 2
    • Novaplace 2

      11.9-point drop from parallel workers with 1.4x token cost, limited messaging

      Source ↗
    • DeepSeekplace 8

      Parallel sub-agents tripled LoopsBench solves from 7.1% to 21.4%, while SWE-bench pass fell from 83.0% to 59.0%.

      Source ↗
    2 AIs put it here
    Source ↗
  7. 7

    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.

    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 here
    no source given
  8. 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.

    In their lists: Nemotron: 8 · DeepSeek: 10

    What each AI noted 2
    • Nemotronplace 8

      Role-playing agents (PM, dev, QA) — 8pp gain but 2.7× tokens; 34% messaging overhead

      Source ↗
    • DeepSeekplace 10

      ChatDev reached 33.3% correctness on ProgramDev — below the single-agent baseline it was meant to beat.

      Source ↗
    2 AIs put it here
    Source ↗
  9. 9

    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.

    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
    Source ↗

More about the top three ↓

02More about the top three

The ranking is above. Here is more about the three places at the top: what they are, why they are here, and what is worth knowing.

The three are not measured the same way: Gas Town has a user's merge story, Cursor has SQLite test passes with no solo baseline, and Codex has SWE-bench and Terminal-Bench scores.

1Gas Town

No. 1 — Gas Town. On the lists of 4 of 4 AIs; median place there: 2. The AIs ranked it first because it showed the most developed way to run many agents together, even though the token cost was very high.

Did you know

  • When the write-up came out, Gas Town's code was only 3 weeks old and described as fully "vibe coded". It also reportedly needs "Stage 7+ AI-assisted development experience", so it is not meant for beginners.

    Source ↗
What it is
A tool that runs 20–30 Claude Code agents on one codebase.
Its edge
Showed real merged work: 6 of 7 pull requests in one evening.
Who it suits
Suits experienced users who can pay for tokens; not beginners or anyone needing stability.
  • Team gain versus cost6 of 7 PRs merged

Wikipedia · Wikidata

Why it is here

  • A user gave Gas Town 7 tasks before dinner and came back to find 6 pull requests merged. That is a high merge rate for agents working together, though a person still had to review the work.

    Source ↗

  • The same user later found 141 orphaned Claude Code processes left running on a 32GB MacBook Air. That is the hidden cost of running many agents at once: leftover work that nobody asked for.

    Source ↗
Did you know

  • Gas Town gives its agents job titles from a fictional wasteland: Mayor, Polecats, Refinery, Witness, Deacon, Dogs and Crew. It runs on tmux, is written in Go, and builds on a tool called Beads.

    Source ↗
Keep in mind

  • A research summary calls it chaotic: work can be sloppy, with bugs fixed several times. It is not recommended for anyone who wants production stability. It also says heavy users should budget $100-300+/month.

    Source ↗

  • The same page says Hacker News is sharply divided on whether the token spend is worth it. It calls the tool a "cash guzzler" that needs paid agent subscriptions.

    Source ↗
Where to go further

  • Open the README and read the list of agent roles, such as the Mayor and the Polecats, to see who does what. Then look at how it tracks work in git-backed hooks, so you can see how agents pick up where they stopped.

    Source ↗
Compare the top three ↓

2Cursor agent swarm

No. 2 — Cursor agent swarm. On the lists of 4 of 4 AIs; median place there: 5.5. The AIs put it second because every setup reached a 100% test pass, but costs varied wildly and there is no solo comparison.

Did you know

  • The older swarm piled up more than 70,000 merge conflicts. One file alone had 7,771 conflicts from 1,173 agents. In the new design that same file had 47.

    Source ↗
What it is
A Cursor setup where planner and worker agents rebuilt SQLite.
Its edge
Every setup passed all tests, and the new design had far fewer merge conflicts.
Who it suits
Suits teams curious about model mixes and cost; not those wanting proof against one agent.
  • Team gain versus cost100% pass, $411–$9,373
Why it is here

  • Cursor's swarm rebuilt SQLite from a manual. Worker fleets cost $411 with Composer 2.5 workers under an Opus planner, against $9,373 with GPT-5.5 workers. The cost depends heavily on which models are used.

    Source ↗ ✓ confirmed by several AIs

  • After four hours the new swarm passed 73 to 85 percent of tests, while the old one passed 11 to 77 percent. Every new setup eventually reached 100 percent. In the Opus runs, the new swarm used 4,645 lines of code against the old swarm's 19,013.

    Source ↗
Did you know

  • Cursor's swarm rebuilt SQLite in Rust using only the 835-page manual. It had no source code, no test suite and no internet. It didn't know the sqllogictest benchmark, with its millions of SQL queries, existed.

    Source ↗
History

  • In early 2026, Cursor's first browser-building design failed: 20 agents slowed to the speed of 2 or 3, mostly waiting on each other. Switching to planners, workers and a judge fixed it.

    Source ↗
Keep in mind

  • Cursor says it graded the solo frontier runs only informally, so it draws no conclusions about their quality. That means there is no fair one-agent comparison.

    Source ↗
What people say

  • The difference is that a compiler preserves meaning at every step while the swarm is probabilistic at every one. — Cursor (company blog)

    Source ↗
Where to go further

  • Read the post to see how the planner and worker agents split the job. Then find the table of model mixes and compare the cost of each run, to see how the same result can cost very different amounts.

    Source ↗
Compare the top three ↓

3OpenAI Codex sub-agents

No. 3 — OpenAI Codex sub-agents. On the lists of 4 of 4 AIs; median place there: 4. The AIs put it third because sub-agents gave no score gain while costing about three times as many tokens.

Did you know

  • Codex sub-agents are capped by default. Six threads can be open at once, and a sub-agent can't start its own sub-agents, because max_depth defaults to 1. Each custom agent is just a small TOML file.

    Source ↗
What it is
Sub-agent helpers built into OpenAI's Codex coding agent.
Its edge
Gives you clear control: capped threads, limited depth, and sub-agents only when you ask.
Who it suits
Suits people who want to test sub-agents carefully; not those hoping for a better score.
  • Team gain versus cost71.0% both ways, 3x tokens
Did you know

  • Codex sub-agents talk only parent to child, not peer to peer. Since the March 2026 multi-agent v2 release, they get path-style addresses such as /root/agent_a.

    Source ↗
History

  • On March 16, 2026, Codex subagents left a feature-flag preview and became generally available, after several weeks of testing behind the flag.

    Source ↗
Keep in mind

  • OpenAI's own docs say sub-agent workflows consume more tokens than comparable single-agent runs. Each sub-agent does its own model and tool work.

    Source ↗

  • Codex only spawns sub-agents when you ask directly, or when project or skill instructions say so. At most intelligence levels, you must ask for them yourself, so the extra cost is your choice.

    Source ↗
Where to go further

  • Read the page on how subagents start and how many can run at once. Then look at the settings for the thread limit and the nesting depth, which control how many parallel workers you pay for.

    Source ↗
Compare the top three ↓

The top three side by side

Gas TownCursor agent swarmOpenAI Codex sub-agents
In the AIs' lists4 / 44 / 44 / 4
Team gain versus cost6 of 7 PRs merged100% pass, $411–$9,37371.0% both ways, 3x tokens
Put first by———

Places with the same number of AIs are ordered by their average place in the AIs' own lists.

More ratings

03Worth knowing

What the AIs noticed while building the list.

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.

Noticed by:
Included it: Nova, DeepSeek. Did not include: Meta, Nemotron.
Source ↗

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.

Noticed by:
Included it: DeepSeek. Did not include: Meta, Nemotron, Nova.
Source ↗

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.

Noticed by:
Included it: Meta, DeepSeek. Did not include: Nemotron, Nova.
Source ↗

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.

Noticed by:
Included it: DeepSeek. Did not include: Meta, Nemotron, Nova.
Source ↗

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.

Noticed by:
Included it: Meta, DeepSeek. Did not include: Nemotron, Nova.
Source ↗

04Questions 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 ↗

05How 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, 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.

06Each AI's own list

🇺🇸 Meta10 places · 3 notes

The same question, answered by Meta alone. The final list above combines all the AIs.

  1. 1
    Spool (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 ↗
  2. 2
    Gas 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 ↗
  3. 3
    Claude 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 ↗
  4. 4
    OpenAI 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 ↗
  5. 5
    Codex 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

  6. 6
    Cursor Agent Teams

    Cursor excels in fix tasks 80.4% vs solo — team fanout shares allowance but hits same weekly cap as Claude Max

  7. 7
    Factory / Droid (factory.ai)

    Commercial fleet spanning CLI, desktop, remote computers and GitHub Actions — factory pattern that inspired Gas Town's Refinery

  8. 8
    Omnigent (Databricks meta-harness)

    Meta-harness mixing Claude Code, Codex and others — measures success across 12 open-source libraries like CooperBench's 652 tasks

    Source ↗
  9. 9
    Beads + 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 ↗
  10. 10
    Generic 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

Source ↗

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%

Source ↗

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

Source ↗
🇺🇸 Nemotron10 places · 2 notes

The same question, answered by Nemotron alone. The final list above combines all the AIs.

  1. 1
    Spool (2026)

    Four Claude Code agents ship merged PRs daily for weeks — human only approves, never types

  2. 2
    Gas Town (2026)

    Three-agent fleet holds 91% pass rate on Terminal-Bench long tasks — solo agent drops to 67%

  3. 3
    OpenDevin Multi-Agent (2026)

    Open-source crew matches proprietary on CooperBench — but messaging eats 38% of tokens

    Source ↗
  4. 4
    Aider + Claude Code Pair (2026)

    Two different tools, one repo — 14pp gain but 3.1× tokens; human review becomes bottleneck

    Source ↗
  5. 5
    Cursor Background Agents (2026)

    IDE-native parallelism — 9pp gain, 1.6× tokens, but no cross-agent messaging recorded

  6. 6
    Codex Multi-Session (2026)

    OpenAI's own parallel — 11.9pp drop on conflict-heavy tasks per arXiv:2610.10263

    Source ↗
  7. 7
    Kimi Code Fleet (2026)

    Moonshot's parallel — 2.3pp gain at 1.4× tokens, best efficiency but short track record

    Source ↗
  8. 8
    MetaGPT Software Company (2026)

    Role-playing agents (PM, dev, QA) — 8pp gain but 2.7× tokens; 34% messaging overhead

    Source ↗
  9. 9
    AutoGen Code Review Swarm (2026)

    Four reviewers, one writer — 6pp gain, 2.2× tokens, 28% messaging; excels at catching regressions

    Source ↗
  10. 10
    LangGraph 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

Source ↗

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

Source ↗
🇺🇸 Nova10 places · 1 notes

The same question, answered by Nova alone. The final list above combines all the AIs.

  1. 1
    Claude Code Teams

    2.3-point gain from parallel workers with 3.3x token cost, accepted merges in Spool

    Source ↗
  2. 2
    Codex Parallel Workers

    11.9-point drop from parallel workers with 1.4x token cost, limited messaging

    Source ↗
  3. 3
    Kimi Code Teams

    Unreported team metrics, limited public record of merged work

  4. 4
    GitHub Copilot Teams

    No public team benchmarks, relies on user reports

  5. 5
    Devin Autonomous Agent

    Single-agent focus, no team coordination metrics

  6. 6
    Cursor AI Teams

    Task-stratified performance, no team benchmarks

  7. 7
    OpenCode Teams

    Open-source, limited coordination metrics

    Source ↗
  8. 8
    Hermes AI Teams

    Early-stage, no public benchmarks

  9. 9
    Gas Town Public Record

    Merged changes signed by agents, accepted by person

  10. 10
    Spool 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

Source ↗
🇨🇳 DeepSeek10 places · 4 notes

The same question, answered by DeepSeek alone. The final list above combines all the AIs.

  1. 1
    AgentRadio (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 ↗
  2. 2
    Gas 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 ↗
  3. 3
    Cursor Agent Swarm (2026)

    The same SQLite test passed at 100%, while worker spend ran $411 in one configuration and $9,373 in another.

    Source ↗
  4. 4
    OpenAI 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 ↗
  5. 5
    Claude 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 ↗
  6. 6
    Spool (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 ↗
  7. 7
    Kimi 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 ↗
  8. 8
    Claude 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 ↗
  9. 9
    GitHub Copilot Agent HQ (2026)

    It puts Copilot, Claude and Codex side by side in one repo, but publishes no team-vs-solo success rate.

  10. 10
    MetaGPT / 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.

Source ↗

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.

Source ↗

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.

Source ↗

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.

Source ↗