A 17-agent product team that senses, prioritises, de-risks, ships, and learns on your real systems, Linear, Notion, Slack, and Amplitude, with deterministic gates that decide what it can do on its own and what needs you.
Built on Anthropic's official product-management plugin. It keeps every skill you already know (write-spec, roadmap-update, stakeholder-update, synthesize-research, competitive-brief, metrics-review) and adds the operating layer: evidence gates, adversarial verifiers, drift detection, and a closed launch loop. See Anthropic's product-management plugin
3 sales calls + 11 deals blockedPassamplitude://funnel/activation -38% step 3Pass142 support tickets clusteredPassamplitude://retention flatEscalatereversible but low confidence(none), requested by CEOBlockno cited evidence, HiPPO rejectedThe shift
Same product moments. The difference is whether the work ends as a recommendation you still have to action, or as a gated action the system performs and records.
It returns a clear answer. You do the rest.
It performs the action, behind a gate, and leaves an evidence trail.
When an action is irreversible, high blast radius, or thin on evidence, it does not act. It stops and asks you.
Install
Install in Claude Code from the marketplace, or upload the Claude Cowork plugin. Both ship the full enforcement spine. You connect your own systems when you are ready to switch on live operation.
Two commands in your terminal. The plugin, its agents, the hooks, and the enforcement spine install together.
claude plugin marketplace add mlobo2012/autonomous-pm-plugin
claude plugin install autonomous-pmFor local development, point Claude Code at the plugin directory with claude --plugin-dir /path/to/autonomous-pm-plugin.
Enter your work email and the Cowork plugin zip will open. Then upload it in Claude Cowork and connect your systems.
No spam, ever. Your email is stored securely so we can send you updates about new use cases and workflows.
The four surfaces
Rendered from the plugin's own logic, not a stock dashboard. Every score, verdict, and assumption carries a gate light: green Pass with cited evidence, amber Escalate when a human is needed, red Block when there is no evidence or the action is irreversible.
3 sales calls + 11 deals blockedPassamplitude://funnel/activation -38% step 3Pass142 support tickets clusteredPassamplitude://retention flatEscalatereversible but low confidence(none), requested by CEOBlockno cited evidence, HiPPO rejectedAt least 30% of enterprise accounts enable SSO within 30 days of GA.
2 of 3 reference calls unsureSCIM raised in 2 calls11 deals tagged sso-blockerStrongest case it flops: buyers want SCIM provisioning, not just SSO. 2 of 3 reference calls raised it.
Cheapest killer test: 5 customer-development calls before any build.
An assumption is only marked contradicted when both the documented claim and the opposing signal are present.
61% of new activations started on mobile (amplitude).ContradictedEnterprise pipeline up 22%, no opposing signal.HoldsSetup completion steady at 74%, no divergence.Holdsbelow pre-registered target. Victory-Bias Auditor blocked the success write-up.
cleared the target. Expansion bet opened with evidence.
partial signal. Adjudicator queued one more cycle.
The control model
Three deterministic gates sit under every action. They are hooks and libraries, not a tone of voice. They decide what is safe to run alone.
No score, verdict, or decision passes without a cited evidence record. hooks/evidence-gate.sh blocks the claim if the citation is missing.
Success is impossible below the pre-registered target. lib/launch.py refuses to call a launch validated when the actual missed the number.
lib/escalation.py classifies each action by reversibility, blast radius, confidence, and agent disagreement, then routes the rare strategic call to a human with a brief.
The team
Each pod owns one part of the loop. The four marked agents are adversarial verifiers, placed where a confident mistake is most expensive.
Adversarial verifier
Built on Anthropic
Autonomous PM inherits the official product-management plugin. Out of the box each skill gives you advice. With the operating layer, each one performs the gated action and leaves an evidence trail.
write-specTurns an idea into a PRD with requirements, scope, and success metrics.Maps the load-bearing assumptions, red-teams the spec, and holds the ship gate until the killer test passes.roadmap-updateHelps refresh or reprioritise a roadmap in a familiar format.Reprioritises from this week's cited signal and moves the Linear cards when the change is reversible.stakeholder-updateDrafts a status update for executives or partners.Posts the update to the right Slack channel with the evidence and the dissent attached.synthesize-researchTurns interviews and tickets into themes and opportunities.Keeps signal flowing always-on and clusters it into cited evidence the gates can use.competitive-briefCreates a competitor brief with positioning and implications.Feeds the brief into drift detection so a competitor move can flag a contradicted assumption.metrics-reviewReviews product metrics, trends, and follow-up actions.Pre-registers the prediction, pulls the actual at the decision date, and adjudicates iterate, hold, double down, or kill.brainstormStress-tests ideas and explores the problem space.Routes the best options into the evidence gate so nothing is prioritised on a hunch.What you are actually getting
This is not zero-setup full autonomy. It is calibrated trust, which is the point.
Most of the autonomous PM category is agent washing. Height, the most hyped pure-play, shut down in 2025. The point was never autonomy for its own sake. It is the discipline underneath: cited evidence, deterministic gates, and a closed loop you can audit. Autonomous PM is the disciplined version, built on Anthropic's plugin rather than a replacement for your stack.
In Claude Code from the marketplace, or in Claude Cowork from the plugin zip. The enforcement spine and the demo cycle run on day one.
Linear or Jira, Notion, Amplitude or Pendo, Slack, and your research and support sources, each with its own login. Until you do, it cannot know your real product world.
You decide how far it can act on its own. Start read-only, widen the scope as you trust it.
Reversible and well-evidenced actions can run alone. Irreversible, high blast radius, or thinly evidenced ones stop and ask you.
Fit
FAQ
Anthropic's plugin gives you the PM skills and returns advice. Autonomous PM keeps those skills and adds an operating layer: a 17-agent team, an evidence substrate, adversarial verifiers, drift detection, a closed launch loop, and deterministic gates that decide when the system may act and when a human must approve.
It can act, behind a gate. When an action is reversible and cited, it can move Linear cards, open a ticket, or post a Slack update on its own. When an action is irreversible, high blast radius, or thin on evidence, it stops and asks you.
A hook, hooks/evidence-gate.sh, that blocks any score, verdict, or decision that does not cite a real evidence record. It is how the system kills HiPPO: nothing is prioritised because a leader asked, only because a metric, ticket cluster, or transcript backs it.
Four agents placed where a confident mistake is most expensive: Evidence Adversary, Pre-Mortem Red Team, Victory-Bias Auditor, and Consensus Checker. They argue against the confident call, name the cheapest test to kill a risk, and block success claims that missed the target.
Your own systems through their MCP servers: Linear or Jira for the roadmap, Notion for knowledge, Amplitude or Pendo for metrics, Slack for broadcasts, and your research and support sources for signal. Without those connections the enforcement spine and demo cycle still run, but live writes and always-on pulls cannot know your real product world.
That is what the human-escalation boundary is for. lib/escalation.py classifies each action by reversibility, blast radius, confidence, and agent disagreement. You start read-only, widen the autonomy scope as you build trust, and the gates hold the line.
Both ship the full enforcement spine. Use Claude Code if you want the marketplace command and the repo visible. Use Claude Cowork if you want the visual plugin upload path. The install section on this page covers both.
Yes. The plugin is free and open source on GitHub. You only pay for the third-party tools you choose to connect, such as your existing Linear, Notion, or Amplitude accounts.
Sources and research
The gates are not opinion. They are built on the product research that shows why most features go unused and why confident calls go wrong.
80% of features are rarely or never used. The build trap, measured.
View source64% of software features are rarely to never used.
View sourceAbout one third of experiments improve the metric they target.
View sourceThe highest-paid person's opinion is what evidence-gated priorities exist to replace.
View sourceTwo of three product ideas fail to move the metric they target.
View sourceNetflix reports about 90% of ideas fail to beat control. Google reports about 96%.
View sourceOnly about 1 in 10 experiments produced a statistically significant win.
View sourceThe MAST taxonomy reports 41% to 86.7% failure rates. It is why the gates are deterministic, not vibes. arXiv:2503.13657.
View sourceOpportunity Solution Trees, the discovery discipline behind the ship gate.
View sourceThe feature factory failure mode the closed loop is built to break.
View sourceCustom build
The free plugin is the clean starting point. If your product org runs across Linear or Jira, Notion, Amplitude or Pendo, Slack, support tickets, sales calls, and rules that live in people's heads, Marco will build the operating loop around the way your team actually works.
Linear or Jira, Notion, Amplitude or Pendo, Slack, support, and the places product work really happens.
Evidence rules, escalation thresholds, the autonomy scope you trust, and the product judgement the agents must respect.
Reprioritisation, ship gates, drift scans, and launch loops inside your workflow, with humans approving the rare strategic call.
Install Autonomous PM
Free and open source. Claude Code marketplace or Claude Cowork plugin. The gates decide what is safe to run alone.
Talk to Marco