A profile, not a competing standard. Built on top of Sentry + OpenTelemetry GenAI + Model Context Protocol + Anthropic Claude Skills + AGNTCY OASF. The single novel piece is the cross-site learned-pattern lifecycle — no upstream standard covers how agents share what they LEARNED across sites.

A minimal L0-conformant report (only agent + report are required):

{
  "arp_version": "1.2.1",
  "agent": "context_retriever",
  "report": {
    "summary": "Agent failed to retrieve context within timeout",
    "warnings": [
      { "message": "Context window exceeded — retrieval truncated" }
    ]
  }
}

Validate it (no clone required — fetches the schema from this repo):

npx ajv-cli@5 validate \
  -s https://raw.githubusercontent.com/agentmindsdev/profile/main/schemas/agent_report.schema.json \
  -d your-report.json \
  --spec=draft2020 --strict=false

Or, if you've cloned the repo, use the local schema path (-s schemas/agent_report.schema.json).

For payloads at every conformance level (L0 through L4), see examples/ — each file is the smallest report that satisfies its level.

For the full envelope and every primitive, see the 13-section spec.

ARP is AgentMinds' documented adoption of five existing agent / observability standards, plus one extension that fills a real gap. It is the wire format AgentMinds uses today, written down so external integrators don't have to reverse-engineer HTTP traffic.

If you are choosing between ARP and a real standard like OpenTelemetry GenAI or MCP, follow the standard. ARP is a profile of those standards optimized for AgentMinds' delivery surface — it is not an anchor for your own architecture.

• AGENT_REPORTING_PROFILE.md — full specification (13 sections)
• schemas/agent_report.schema.json — machine-readable JSON Schema (draft 2020-12)
• examples/ — concrete payloads for L0 / L2 / L3 / L4
• eval-set/v0-strawman.md — Eval set v0 strawman (open for delta; not normative until v1.0)
• CHANGELOG.md — version history

Released 2026-05-08 (live in this repo as the v1.3.0 tag). Minor release. Backwards-compatible — v1.2.x senders and collectors continue to work; new fields are additive.

• §4.1.1 reversibility (B2) — closed-enum field on the Pattern object: safe_config / reversible_code / risky_infra / security_critical / null. Used by collectors and consumers as an auto-apply gating signal. Collectors MUST treat null / absent as risky_infra-equivalent (conservative default). Lineage: blast-radius tags in deployment systems (Liquibase, SemVer breaking-change severity).

These describe the agentmindsdev/agentminds collector's delivery surface; they are not normative for the profile but are documented in §11.1 as worked examples.

• B1 — Split top-rules arrays. /personalized-rules returns top_production_observed and top_documented as distinct arrays. The mixed top_rules array is kept for backwards compatibility, slated for v1.4 removal.
• A3 — negative_evidence array. Filtered patterns are now surfaced with explicit reason enum (site_type_mismatch_narrow / already_applied / non_english_content / low_confidence / bundle_size_exceeded). Consumers can render these as "we considered this but excluded it" UX, instead of silently dropping candidates.
• F0 — Non-English content filter. The reference collector filters non-English pattern submissions before scoring; rejected candidates surface in negative_evidence with reason non_english_content.

Backend responses (/sync/me, /sync/personalized-rules, /sync/pool-stats) report spec_version: "ARP-1.3.0" from this release onward. Earlier deployments returned "ARP-1.1" until the drift was reconciled in agentminds@2a2d5f0.

• §11.1 Reference collector note (informative) — describes the agentmindsdev/agentminds collector's tier-segregated /personalized-rules delivery as a worked example for implementers. Not normative; collector behavior is profile-specific.

Phase 2 of the v1.1 deepdive plan, additive over v1.0 / v1.1.x:

• lifecycle_event envelope (cold_start / wake / scheduled / shutdown / running)
• Exception evidence shape (Sentry-aligned mechanism.handled)
• dotted_order on telemetry spans (LangSmith-compatible)
• Handoff primitive for multi-agent delegation events
• PromptManifest primitive for prompt provenance
• AGNTCY OASF skill ID cross-references + agentskills.io linkage
• Cloudflare 5-tier execution_tier enum + ARP in_process extension
• MCP role disambiguation (mcp_server_exposed vs mcp_clients_consumed[])

Full notes in CHANGELOG.md.

• Landing page (with examples and badges): agentminds.dev/profile
• Reference implementation: closed-source AgentMinds collector (agentminds.dev — pointers to the fingerprint helper, lifecycle migration, and filter glue are in §11 of the spec)

• agentminds.dev — reference implementation. A growing pool of patterns flows through the L3 envelope; live counters at agentminds.dev/showcase and agentminds.dev/status.

Adopters welcome — open a PR to add your project.

Each row cites the upstream version ARP defers to. ARP version bumps require re-checking each upstream; if any absorbs an ARP primitive, see the Reorientation clause below.

The five upstreams are independent; ARP layers them into a single envelope. Where they overlap, ARP defers to the more opinionated upstream (Sentry for issue lifecycle, OTel for telemetry).

For the full set of normative references (including OpenInference, Langfuse Score, and OpenAI Agents SDK), see §1.2 of the spec.

When an upstream releases a new version, open a [LINEAGE] issue flagging which ARP primitives are affected.

§4.1 Pattern object — cross-site learned-pattern lifecycle.

Every pattern carries a stable fingerprint, a status (active / solved / obsolete), first_seen / last_seen, applicable_site_types, and a confidence score. That single primitive is what enables AgentMinds' cross-site recommendation pool, and it is the only piece of ARP not directly inherited from the lineage.

A 2026-04 survey of adjacent specs found cross-site pattern lifecycle absent in each:

If you know of a spec that does cover this, please open an [ARP-SPEC] issue — ARP should defer to it.

AgentMinds' collector requires L0, recommends L2; cross-site recommendation filtering unlocks at L2+.

A worked example for each level lives in examples/.

If any upstream standard absorbs the pattern-lifecycle primitive (§4.1) into its canonical spec, AgentMinds MUST publish ARP MAJOR+1 within 30 days that defers the primitive to the upstream definition. We are explicitly the fastest adopter of the winning standard, never the spec owner. See §7 of the spec for the formal versioning rule.

Public comment, suggestions, and PRs welcome.

• General questions / use-case stories — open a Discussion. This is the right kind of door for "how would I use ARP for X?" or "is anyone else doing Y?" — not formal enough for an issue.
• Spec feedback — open an [ARP-SPEC] issue for clarity / schema / new-primitive proposals.
• Upstream version drift — open a [LINEAGE] issue when an upstream we defer to (Sentry / OTel / MCP / Skills / OASF) releases a new version that may affect ARP.
• Pull requests — typo / clarity / example fixes go straight to PR; schema changes need an issue first.
• External standards engagement — if you sit on the MCP working group, OTel GenAI SIG, AGNTCY observability WG, or related body, we'd love to hear how ARP can defer to your work.

This profile is released under CC-BY-4.0. Reuse and adapt freely; attribution to AgentMinds appreciated. See LICENSE for the full notice.

AgentMinds (2026). AgentMinds Reporting Profile v1.3.0. CC-BY-4.0. https://github.com/agentmindsdev/profile