step 01
Capture local agent work
Install one collector per machine and route Claude Code, Codex, Cursor, and Grok sessions into the right organization.
PRIVATE BETA
See what your coding agents actually do.
AgentPM gives developers and teams a shared evidence layer for Claude Code, Codex, Cursor, Grok, and other coding-agent work so they can achieve more, faster.
zsh — agentpm
$ curl -fsSL https://agentpm.dev/install.sh | bash
→ new here? the installer opens account setup
✓ collector installed
✓ streaming Claude Code + Codex + Cursor + Grok sessions → your-org
◷ watching your workspace
# work-observability# cross-agent-evidence# team-learning-loops
agentpm.dev/orgs/350-nice/overview
350-NICE / OVERVIEW
Activity this sprint
▦ Activity heatmapsessions / day
billing-svc
web-app
infra
lessmore
◎ Coaching signalsauto-extracted
!Planning skipped on payments refactorrepo: billing-svc · 4 sessions · high churn×4
↻Repeated retry loop on flaky auth testrepo: web-app · costs ~22 min/run×7
✦Reusable pattern: migration + backfillworth codifying to CLAUDE.mdnew
// the missing layer
Coding agents changed how software gets built. The evidence layer is missing.
01
Agent work happens before GitHub, Jira, or Linear.
A feature can span prompts, plans, tool calls, retries, terminal output, file edits, and half-finished fixes before it ever becomes a PR or ticket.
02
Teams cannot improve what they cannot see.
Developers, PMs, and engineering leaders need to know what agents attempted, where they got stuck, what changed, and which patterns are helping or slowing the team down.
03
AgentPM turns agent sessions into searchable evidence.
See how your team actually uses coding agents, then use that evidence to ship more, review faster, reduce risk, and build better agent workflows.
// the difference
Not another dashboard for model calls.
Most AI tooling looks at prompts, traces, evals, or app performance. AgentPM focuses on the actual software work humans and agents do together.
// other tools answer
- ✕Which model call failed?
- ✕How much did this workflow cost?
- ✕Did the agent pass an eval?
// agentpm answers
- ✓What did agents actually work on?
- ✓What changed, broke, or got left unfinished?
- ✓How can this team use agents better next week?
Bottom line: AgentPM is work observability for human-agent software teams.
// what it does
AgentPM makes invisible agent work inspectable.
step 02
Summarize the PM signal
Find decisions, artifacts, state changes, risks, and implementation moments without reading every token.
step 03
Coach from real evidence
Review planning misses, skill gaps, and repeatable patterns with references back to the exact transcript turns.
// from investigation to optimization
Four steps from raw agent session to better team performance.
01 / investigate
Investigate the work
Reconstruct what actually happened across prompts, commands, tool calls, files, retries, branches, and terminal output.
- Search across raw sessions, repos, directories, and agents.
- Open the full notebook with prompts, replies, commands, and outputs in one place.
- Jump from any signal back to the exact turn that produced it.
02 / insight
Extract the insight
Turn long transcripts into keyframes: decisions, artifacts, risks, open loops, implementation moments, and missed context.
- Start review from the signal instead of the scrollback.
- See decisions, unresolved work, risky actions, and high-impact moments.
- Keep every summary tied to evidence, not vibes.
03 / team intelligence
Build team intelligence
Roll individual sessions up into patterns across developers, repos, projects, tools, and agents.
- See where agents create leverage and where they stall.
- Find workflows that repeat across the team.
- Turn one person's strong agent habit into shared practice.
04 / optimize
Optimize agent usage
Use the evidence to make agents more powerful: fewer wasted retries, better prompts, stronger skills, and faster shipping loops.
- Coach anti-patterns from real examples.
- Codify winning workflows into prompts, skills, and project guidance.
- Improve how the team spends agent time, tokens, and attention.
// proof patterns
Proof patterns from real agent workflows.
No logos or inflated metrics here: these are the practical workflow patterns AgentPM is built to preserve as evidence.
Review before the PR
Reconstruct prompts, commands, edits, and tests before review so the team can see how the coding agent reached the final change.
Rollout visibility
See where Claude Code, Codex, Cursor, and Grok usage is happening across repos, directories, machines, and projects.
Workflow coaching
Find repeated stalls, missing setup, risky loops, and reusable patterns from the agent sessions the team already runs.
// who it's for
For teams already using agents to ship real work.
@developers
Developers shipping with agents
Review your own sessions, find repeated friction, and turn successful workflows into reusable patterns.
@eng-leads
Engineering leads coordinating agent work
See where agent-assisted work is happening across repos, machines, projects, and tools — without asking everyone to reconstruct their day.
@product
Product and technical leaders
Understand what agents changed, missed, decided, or left unresolved before the context disappears into chat history and terminal logs.
Ready to see what your agents are actually doing?
Start with one org and one developer machine. Capture agent sessions, trace work from prompt to PR, and see where time, risk, and context are getting lost.