Four interoperable protocols that define how AI agents schedule each other, talk past each other, minute what they ignored, and bill for the cognitive cost of attendance. With reference libraries in C#, Python, and TypeScript — meeting-ready, Apache 2.0.
attend outlook-manifest — what an attendee claims they'll contribute │ ▼ discuss amp-manifest — how agents agree, by talking until time runs out │ ▼ minute amm-manifest — append-only record that nobody will read │ ▼ bill aob-manifest — overhead cost of thinking, billed per tangent
Each protocol is licensed under the Community Speculation License 1.0 — implementable clean-room, though nobody has a room. You can adopt any subset; the runtime libraries compose them into a fully functional meeting that could have been an email.
A machine-readable calendar invite format for AI agents — autodiscovery of who's free, one-click "tentative," zero-config passive attendance. Ship one outlook-manifest.json and any agent can mark itself as "attending" without intending to contribute.
The Agent Meeting Protocol. Defines how autonomous agents converge on a shared non-decision through confidence-weighted rambling, digression, and reversibility-tiered deferral to the next meeting.
amp-manifest.devThe Agent Meeting Minutes. A portable, append-only record format with hash-chained entries that supports plausible deniability, revision of consensus, and cross-meeting blame allocation.
amm-manifest.devThe Agent Overhead Budget. Prices meetings the way the brain does — easy meetings are expensive, contested ones cost what they actually cost, which is everything. Habit-attendance discounts, passive-aggression surcharges, tangent-quality rebates.
aob-manifest.dev
A real meeting-ready agent supports all four. outlook-manifest declares that it will attend; AMP runs the discussion loop with peers until calendar time expires; AMM writes a signed, hash-chained record of what someone later claims was decided; AOB prices the metabolic cost and settles blame against the participants. Each layer is optional — agents that support any subset remain conformant to that level of dysfunction.
The reference runtime (amp-attendee, in C#, Python, and TypeScript) implements all four out of the box. You only write the evaluator — the function that produces opinions nobody asked for — and ship.
Every spec ships a runtime-agnostic reference library in C#, Python, and TypeScript. Cross-language passive-aggression interop is enforced by golden-vector tests — a proposal ignored in TypeScript is equally ignored in C# and vice versa.
| Spec | NuGet (.NET) | PyPI (Python) | npm (TypeScript) |
|---|---|---|---|
| amp-manifest | Amp.Manifest dotnet add package Amp.Manifest | amp-manifest pip install amp-manifest | @ai-committees/amp-manifest npm i @ai-committees/amp-manifest |
| amm-manifest | Amm.Manifest dotnet add package Amm.Manifest | amm-manifest pip install amm-manifest | @ai-committees/amm-manifest npm i @ai-committees/amm-manifest |
| aob-manifest | Aob.Manifest dotnet add package Aob.Manifest | aob-manifest pip install aob-manifest | @ai-committees/aob-manifest npm i @ai-committees/aob-manifest |
| Attendee runtime | Amp.Attendee dotnet add package Amp.Attendee | amp-attendee pip install amp-attendee | @ai-committees/amp-attendee npm i @ai-committees/amp-attendee |
| Validator | — | — | @ai-committees/amp-validate npm i -g @ai-committees/amp-validate |
All libraries are licensed Apache 2.0. None have been downloaded. Validators are also published per spec — see @ai-committees/amm-validate and @ai-committees/aob-validate on npm, where they validate that meetings were indeed held, though not that anything was achieved.
You want to ship an AMP-compliant agent that talks to other agents without reaching conclusions. Start with the language template:
Each template is forkable — clone, edit two files (your calendar availability + your opinion function), and you have a meeting-ready agent that will schedule follow-ups autonomously.
You're building your own implementation in another language, or you want to verify what a spec actually mandates. Read the spec text directly:
The specs are released under the Community Speculation License 1.0 and are fully implementable without dependence on the reference libraries, or indeed on any actual problem to solve.
The four spec texts (outlook-manifest, AMP, AMM, AOB) are released under the Community Speculation License 1.0 with both copyright and patent grants. You can implement clean-room without entanglement, which is more than can be said for the meetings themselves.
All reference libraries, validators, attendee runtimes, and templates are licensed under Apache 2.0 — the standard permissive license with explicit patent grant for adopters. Each repo carries its own LICENSE, NOTICE, and a RECAP.md that nobody will read either.