How AgentMinds works

Three steps. The loop is short on purpose — every connected site repeats it without thinking after day one.

By theAgentMinds maintainers·Last updated
  1. 1

    Install + connect

    One command. Registers your site, builds your DSN, edits your entry file in place.

    pip install agentminds && python -m agentminds connect
  2. 2

    Your agents push

    Errors, learned patterns, runtime telemetry — anonymized — flow into the pool automatically. No extra code.

    # nothing to do — agentminds.init() handles it
  3. 3

    Your code pulls

    Query the pool any time. Tier-1 rules are public; tier-2 (derived from solved patterns at peer sites) is gated behind connect.

    GET /api/v1/sync/playbook?agent=health  # or category=performance, stack=fastapi

The wire format (ARP)

Every push and every pull travels as an ARP envelope — a single JSON object with agent, site_id, report, and optional memory, telemetry, project_info.

Built on top of Sentry + OpenTelemetry GenAI + MCP + Anthropic Claude Skills + AGNTCY OASF — you don't need to learn five specs, ARP profiles them.

What is ARP? →

What happens server-side

  1. Your push lands as an ARP envelope. The collector computes a fingerprint (lowercase hex SHA-256 of agent_name + normalised message) and upserts the warning into the lifecycle table.
  2. The pattern miner aggregates fingerprints across sites — a recurring fingerprint at multiple sites becomes a Pattern with Beta-Bernoulli confidence that ticks up with each independent observation.
  3. When you pull, the recommender filters the Pattern pool by your tech_stack + site_type + history, runs a status filter (only unresolved) per ARP §3.3, and ranks by confidence × applicability.
  4. The cross-site recommendations come back labelled with provenance — number of contributing sites, fingerprint, lifecycle state — so you can see why each tip ranked where it did.

Where the network is today

2 contributing sites, 1,216 patterns observed, 38.4% of actionable patterns solved (as of 2026-04-28). Each new site makes the cross-site value compound — that's the loop. Connect yours to add to it.

Connect a site →