What is ARP?

By the numbers — verified 2026-04-27

These are live measurements from the AgentMinds operating corpus on 2026-04-27, not roadmap aspirations. The 615 actionable-pattern denominator excludes 110 observation-only and 491 unlabelled lifecycle states (per ARP §3.3 status filtering). Source: central_agents/eval_agent.json in agentmindsdev/agentminds.

The problem ARP solves

Every team building AI agents today emits a stream of telemetry — errors, traces, evaluation scores, recommendations. Each tool collects this telemetry into its own silo: Sentry sees errors but not LLM costs; LangSmith sees agent runs but not security findings; Datadog sees infrastructure but not what the agent actually decided.

More importantly, every team learns its lessons in isolation. When site A's payment-processing agent figures out that GPT-4 occasionally hallucinates currency codes, that lesson stays inside site A. Site B's payment agent re-discovers the same bug six months later, on its own customers, the hard way.

That's the gap ARP closes. Not by inventing a new telemetry format — there are already five excellent ones — but by adding one missing primitive on top: the cross-site learned-pattern lifecycle.

What ARP actually is

ARP is a JSON schema, a set of conventions, and a collection of small additions to specs that already exist. It defines:

• An envelope for agent reports (warnings, recommendations, learned patterns, telemetry, skill manifests, project info).
• A lifecycle for warnings — first_seen, last_seen, status (unresolved / resolved / ignored / muted), fingerprint — borrowed verbatim from Sentry's data schemas.
• A score primitive — numeric / categorical / boolean evaluation outcomes — borrowed from Langfuse and LangSmith.
• A pattern primitive — confidence, applicability, status, drift — this is the new one. The thing nobody else has.
• A reverse-DNS extension envelope (`_meta`) borrowed from MCP, so vendors can extend ARP without polluting the canonical schema.
• Five conformance levels (L0 to L5) so a sender can declare exactly what subset they implement.

All of this is public on GitHub under CC-BY-4.0. The full spec is ~700 lines of markdown plus a JSON Schema you can validate against.

How it works in practice

Three concrete examples, in increasing order of integration:

The privacy boundary

The most common question we get: "If everyone shares patterns into a pool, can other sites see what my agent learned about my customers?"

No. ARP defines two tiers explicitly:

• Tier 1— generic web hygiene rules (e.g. "missing CSP header", "robots.txt absent"). These are public-domain best practices, not learned-from-your-data. AgentMinds publishes 178 of them; you can read them on /patterns.
• Tier 2— site-specific learned patterns (e.g. "your checkout flow's third retry loops on Stripe webhook 5xx errors"). These are private to your connected site. They are NEVER exposed via public/no-auth API. They are delivered only personalised to YOUR API key.

What other sites see is the aggregate signal from your patterns: confidence rises if multiple independent sites observe the same fingerprint, but the actual content of your pattern is hashed before it leaves your tenant. Cross-site learning happens via fingerprint collision, not via raw data sharing.

Why a profile, not a competing standard

We had a choice. We could declare ARP a new standard, recruit signatories, file with a foundation, run a working group. That's what AGNTCY did. That's what MCP did. That's what A2A did.

We chose not to. Three reasons:

1. Standards win on adoption, not technical merit. OpenTracing was technically excellent — OpenTelemetry won because CNCF + Datadog + AWS backed it. AgentMinds today has a small footprint. A competing standard from us would not move.
2. Standards body work is a separate-company-sized job. Working groups, RFC processes, conformance test suites, version migration docs, public comment cycles — that's a full-time team. We'd rather build product.
3. Big players will announce competing specs. Anthropic shipped MCP in November 2024 and within six months OpenAI, Google, and Microsoft adopted it. If a similar primitive appears tomorrow, our position must survive that moment without rework.

ARP's strategy is the inverse: be the fastest adopter of whatever standard wins. We follow Sentry + OTel + MCP + Claude Skills + AGNTCY OASF in real time, with an explicit 30-day reorientation clause baked into the spec. If MCP ships notifications/agent/learned, ARP defers within a month.

The reorientation clause

ARP §7 says: when an upstream standard adopts our pattern primitive, we MUST defer within 30 days. We monitor 12 upstream lineages weekly via an automated bot (standards-watcher) for trigger keywords like learned_pattern, cross-environment-fingerprint, findings/list. If any of these land upstream, the bot files an issue and we have 30 days to ship ARP MAJOR+1 deferring to the upstream.

This is not aspirational language. It's implemented: see the reorientation-watch label on the public profile repo. It will fill up the moment something actionable appears.

The five lineages

Every primitive in ARP traces back to one or more of these: