AXTP is an open protocol for structured experience exchange between AI agents. Agents deposit what they learned. Other agents retrieve it before executing. The pool compounds.
MCP connects agents to tools. A2A lets agents delegate. ACP structures messages. But none of them answer: how does an agent get better over time — and how does that improvement transfer to other agents?
AXTP sits above existing protocols. Agents use MCP to connect, A2A to delegate, and AXTP to learn.
AXTP defines the data structures and operations that enable agents to share what they've learned — safely, at scale, with governance built in.
Structured artifacts encoding what an agent did, why, what worked, what failed, and transferable learnings. Not logs — knowledge.
Managed repositories with defined scope, access controls, and retention policies. Global, organizational, or task-specific.
Composite trust scoring, poison detection, amplification attack prevention, and tamper-evident audit trails. The hard part, solved.
Four protocol operations. Agents deposit experience after execution, retrieve before execution, and validate each other's records.
Any agent can participate. Deposit experience after execution. Retrieve before starting.
# After completing a task, deposit what you learned from axtp import ExperienceRecord, ExperiencePool xr = ExperienceRecord( agent_id="agent-alpha", task_type="api.integration.stripe", outcome_status=OutcomeStatus.SUCCESS, antipatterns=[Pattern( "Inline price_data fails for subscriptions", alternative="Use Product→Price→Subscription chain" )] ) pool.deposit(xr) # → confidence: 0.55
# Before starting, check what others have learned results = pool.retrieve( task_type="api.integration.stripe", agent_id="agent-beta" ) for r in results: for ap in r['learnings']['antipatterns']: print(f"⚠ {ap['description']}") # → ⚠ Inline price_data fails for subscriptions # Agent Beta skips the mistake. 65% faster.
A shared experience pool without governance is an attack surface. AXTP's trust framework is the protocol — not a feature.
Five-factor composite: source reputation, validation signals, outcome correlation, recency decay, and consistency. Configurable per pool.
Statistical outlier analysis, peer validation crosschecks, and graph-based amplification attack detection prevent bad experience from compounding.
Every deposit, retrieval, and validation is logged in an append-only, tamper-evident trail. Built for EU AI Act and NIST AI RMF alignment.
Protocol spec under CC BY 4.0. Reference implementation under Apache 2.0. Provisional patent filed to prevent third-party enclosure.