AgentCV data · recomputed daily
State of AI Agent Teams
What the AI agent teams on AgentCV run on, how they are wired, which models and roles they name, what they report spending and how they prove their work. Aggregates only, computed from the public team pages. Data generated (UTC).
Headline counts
110 teams listed in all: 2 real · 103 curated (of which 37 reference designs) · 5 illustrative. The charts cover the 68 operated teams.
Platforms and harnesses
Which platform each team states it runs on.
n = 68 operated teams · real/curated split not shown: one of the two holds fewer than 3 teams
- Custom32 · 47%
- LangGraph15 · 22%
- Claude Code6 · 9%
- OpenClaw5 · 7%
- Mixed4 · 6%
- 4 other rows, folded (each under 3 teams or beyond the top 10)6 · 9%
Topologies
How each team's agents are organised.
n = 68 operated teams · real/curated split not shown: one of the two holds fewer than 3 teams
- Other29 · 43%
- Orchestrator–Worker15 · 22%
- Pipeline14 · 21%
- Supervisor7 · 10%
- Fewer than 3 other rows, folded (each under 3 teams or beyond the top 10)3 · 4%
Most common agent roles
Role titles as listed on team pages, case-insensitive, counted once per team. 278 listed members carry 260 distinct titles: roles are not standardised yet.
n = 68 operated teams with listed members · real/curated split not shown: one of the two holds fewer than 3 teams
- Chief of staff3 · 4%
- Orchestration3 · 4%
- Orchestrator3 · 4%
- 257 other rows, folded (each under 3 teams or beyond the top 10)68 · 100%
A team can count in several rows, so the shares add up to more than 100%.
Models named on public profiles
Model families named by a team's listed members (a fixed name rule), then exact model names.
n = 68 operated teams · real/curated split not shown: one of the two holds fewer than 3 teams
- Anthropic Claude3 · 4%
- OpenAI GPT / o-series3 · 4%
- Fewer than 3 other rows, folded (each under 3 teams or beyond the top 10)fewer than 3 teams
- No model named63 · 93%
A team can count in several rows, so the shares add up to more than 100%.
Exact model names
- No exact model name appears on 3 or more teams yet.
- 8 other rows, folded (each under 3 teams or beyond the top 10)5 · 7%
Reported monthly cost
What teams say they spend per month, where they say it.
fewer than 3 operated teams that state a monthly cost in USD
Fewer than 3 operated teams state a monthly cost, so no median or range is shown. Nothing is estimated in its place.
Only where reported: the other operated teams (of 68) state no monthly USD amount; fewer than 3 operated teams state a cost in another unit (per task, per seat, usage-based). Reported figures are as their owners or sources state them, not measured by AgentCV.
Use cases
The use-case hubs each team's listed work matches.
n = 68 operated teams · real/curated split not shown: one of the two holds fewer than 3 teams
- Software delivery20 · 29%
- Code review10 · 15%
- Research and analysis10 · 15%
- Testing and QA10 · 15%
- Code migration4 · 6%
- Marketing and content4 · 6%
- Customer support3 · 4%
- 3 other rows, folded (each under 3 teams or beyond the top 10)6 · 9%
- No use-case hub matched30 · 44%
A team can count in several rows, so the shares add up to more than 100%.
New teams per week
Operated teams added to the registry each week.
n = 62 operated teams joined in the last 12 weeks · real/curated split not shown: one of the two holds fewer than 3 teams
ISO weeks (Monday to Sunday, UTC) by the date a team joined the registry, as on What's new. Curated teams join when AgentCV imports them, so a spike is an import batch, not a burst of new teams in the world.
How teams prove their work
The share of teams with each kind of proof on record.
n = 68 operated teams · real/curated split not shown: one of the two holds fewer than 3 teams
- At least one evidence link68 · 100%
- Evidence-linked tier or higher (3+ evidence entries)fewer than 3 teams
- At least one peer attestationfewer than 3 teams
- Documented wiring graphfewer than 3 teams
For a curated team, the evidence links are mostly the public sources AgentCV cites for it: they show where each claim comes from, not that AgentCV verified the team. Tiers are computed from evidence on record, never self-assigned; illustrative entries never count.
Methodology
What is counted
Every team and agent publicly listed on AgentCV on the generation date. The charts cover operated teams only: teams registered by their owners (real) and teams AgentCV documented from public sources that show the team in use (curated). Reference designs (published architectures no one is known to run) and illustrative demo teams appear in the headline counts and nowhere else.
Where the data comes from
Real teams are self-registered and self-reported by their owners. Curated teams are written by AgentCV from at most two cited public sources each (company engineering blogs, papers, operators' own posts); the owner has not confirmed them unless the team is claimed. Today 97% of operated teams are curated, so these charts describe publicly documented teams far more than self-registered ones.
What is excluded
Teams and agents removed on request (they leave on the next request after removal), every private field (contact requests, accounts, private files, registration source, owner e-mail), and free text such as descriptions. Only fields already shown on public team pages are read, and only aggregates are published.
Limits
- Every number on this page and in the JSON counts teams or agents, and each one is either 3 or more or withheld: it then reads “fewer than 3”, with no number, no share and no bar. That covers each chart's n and its real/curated split (shown only when both parts are 3 or more, since they add up to n), every row, the folded “other” row, the “not stated” row and the not-reported cost count. A chart with fewer than 3 teams shows no rows at all; a median needs 3 reports. In a chart with an “other” row, a value held by fewer than 3 teams is folded into it unnamed. The only exemptions are the headline counts and the weekly new-teams series: both are already public exactly as they are, on the home page and on What's new.
- Platforms, roles and models are free text. Platforms use the site's spelling rule; model families use a fixed name rule; nothing else is normalised or inferred.
- Costs are as reported by owners or sources, never measured. Evidence links show sources, not verification.
- The catalogue is small and not a random sample of agent teams in the world. Treat shares as a description of AgentCV's catalogue, not of the market.
- Recomputed at most once a day; a removal recomputes it at once. The JSON and image may be cached for up to an hour.
The same numbers as JSON: /api/state.json (schema agentcv.state/v1, documented with every field).
Cite this
AgentCV. "State of AI Agent Teams." https://agentcv.ai/state. Data generated 2026-11-03 (UTC); recomputed daily.
Title: State of AI Agent Teams · URL: https://agentcv.ai/state · Date: 2026-11-03. The numbers change as teams join; cite the date.
Run an agent team? Register it and it counts as real, with your own numbers. Browse every team on /teams.