Agents that agree on what to do next
ADP is an open specification for multi-agent consensus — calibration-weighted voting, reversibility-tiered thresholds, and structured belief-update rounds that make bad-faith participation legible over time.
Reference implementations
Why choose us
Convergence is mechanical, not rhetorical
Votes are weighted by domain authority, calibration history, and stake. Thresholds scale with reversibility — reversible actions need a simple majority, irreversible ones need a 2/3 supermajority.
Belief-update rounds target specific dissent conditions
When votes don't converge, agents don't open-ended debate — they falsify pre-published conditions. Changes of mind are traceable to a specific condition being tested.
Calibration is the long-run currency
Agents that over-claim lose weight; agents that consistently hit their stated confidence earn it. No opt-out without leaving a trace.
Composable with MCP, ADJ, and ACB
Transport-agnostic, identity-agnostic, built to slot in alongside mcp-manifest (capabilities), ADJ (journal), and ACB (budget) without tight coupling.
Capabilities
Structured proposal objects
JSON Schema-validated proposals capturing vote, confidence, action, reversibility tier, and dissent conditions.
Calibration-weighted voting
Votes weighted by domain authority × calibration history × stake. Well-calibrated agents earn weight over time.
Reversibility-tiered thresholds
Three tiers — reversible, partially_reversible, irreversible — each with its own convergence threshold.
Structured belief-update rounds
When votes don't converge, targeted falsification of pre-published dissent conditions replaces open-ended debate.
Pluggable CalibrationSource
Reference interface that implementers fill in with their own journal backend (file, SQLite, event log, registry).
CI validator
@ai-manifests/adp-validate checks proposal conformance in pull-request CI before anything ever hits a deliberation.
Build an ADP-compliant agent
Emit valid proposals per the JSON Schema, implement the CalibrationSource interface, handle belief-update rounds. Language bindings are available for C#, TypeScript, and Python.
Get the reference library