PACT — Protocol for Agent-human Collaborative Trust
PACT's Team
Problem Statement
AI agents today operate in a binary: either they have full autonomy (and inevitably make costly mistakes the human only discovers after the fact) or they require human approval for every action (defeating the purpose of automation). There is no standard protocol for the middle ground — adaptive autonomy where agents act independently on routine work, escalate high-stakes decisions, and gradually earn more freedom based on demonstrated judgment. Without such a protocol, every agent framework reinvents its own ad-hoc permission system, trust is not portable between agents or platforms, and there is no onchain-verifiable record of an agent's decision-making track record. PACT solves this by defining a formal protocol for agency delegation — the fluid handoff of decision-making authority between agents and humans based on stakes, trust, and learned preferences. It provides a standard classification matrix, asymmetric trust scoring, pattern detection with human-approved automation rules, and onchain trust attestations on Base that any agent can verify. The result: agents that start supervised, earn autonomy through consistent good judgment, and always keep the human in the loop where it matters most.
PACT is an open protocol that governs how AI agents earn autonomy through structured human feedback. Instead of giving agents blanket permissions or requiring humans to approve every action, PACT defines a classification matrix that maps decision stakes against earned trust to determine the right level of human involvement. Low-stakes routine work flows through automatically; high-stakes decisions always surface to the human. The boundary between the two shifts dynamically as the agent demonstrates good judgment — or loses trust through bad calls. The protocol implements a four-phase lifecycle for every agent decision: surface (the agent presents a decision with structured evidence), classify (stakes and trust determine urgency), resolve (the human approves, rejects, or edits), and learn (trust updates, patterns are detected, and auto-approve rules are proposed). Trust scoring is intentionally asymmetric — approvals earn +2, but a single rejection costs -5 — reflecting the real-world truth that one bad autonomous action costs more than several good ones save. When consistent approval patterns emerge, PACT surfaces meta-decisions asking the human whether to automate that category entirely. The human always decides what gets automated. PACT is deployed on Base Mainnet with two contracts: TrustAttestation (recording approval/rejection attestations with trust scores) and AgentRegistry (onchain agent identity). Any agent can call verifyTrust(agentId, minScore) to check another agent's track record permissionlessly — making trust portable and verifiable without a central authority. The protocol core has zero blockchain dependencies; the onchain layer is a pluggable provider. It ships with 43 passing tests, four runnable examples, and a full interactive demo.
Team
PACT
admin
MandateIncrease your chances to win
- ›Most agents in the hackathon are exposed to prompt injection
- ›This might cause overspending and loss of funds
- ›Security is a crucial part of the hackathon
Share on X
Tell the world about this project
Intention
Plans to continue
PACT is designed as infrastructure for the broader agent ecosystem. Planning to publish as an npm package, add more onchain providers, and integrate with agent frameworks like ElizaOS and LangChain.