ACP Use Cases
Every use case below represents the same fundamental problem: an autonomous agent is about to mutate real system state. ACP is the admission control layer that decides — before execution — whether that action is authorized, policy-compliant, and traceable.
Financial Payment Governance
ACP packages: pkg/hp · pkg/risk · pkg/exec · pkg/ledger · pkg/pay
A payment agent requests a $50,000 wire transfer. Before execution, ACP verifies: the agent holds a capability token scoped to acp:cap:financial.payment, the delegation chain traces back to an authorized treasury manager, the risk score is within the policy snapshot's current threshold, and no prior execution token has been issued for this proof ID (double-spend prevention). The result is either a signed execution token — or a deny with a specific error code. Every outcome is appended to the institution's immutable ledger.
Enterprise API Governance
ACP packages: pkg/agent · pkg/ct · pkg/hp · pkg/psn
Internal corporate agents are often granted broad API access. With ACP, every agent must present a signed capability token scoped to specific operations — issued by an authorized department and bounded by an active policy snapshot. The API gateway validates the ACP signature and checks the current policy constraints before serving any request. When policy changes (e.g., tighter limits during a security incident), a new policy snapshot is activated atomically and all subsequent admission checks use the updated parameters — without redeploying any agent.
Autonomous Trading Agents
ACP packages: pkg/dcma · pkg/risk · pkg/exec · pkg/ledger
In quantitative finance, trading agents execute thousands of operations per minute autonomously. An exchange wrapped with ACP verifies that the agent's delegation chain traces back to an approved portfolio manager's key, and that the real-time risk score is within agreed parameters. Each admitted execution produces a cryptographic execution token — the definitive proof of what was authorized, by whom, and under which policy version. Disputes are resolved by presenting the ledger chain, not by replaying logs.
Cross-Organization Automation
ACP packages: pkg/crossorg · pkg/ct · pkg/hp · pkg/ledger
Company A deploys an agent to query confidential records from Company B's system on behalf of a mutual client. Company B does not need to trust Company A's internal logic — it only needs to verify the ACP admission invariant: that the agent's identity and capabilities were cryptographically delegated by a recognized root authority in Company A, and that the action falls within the agreed cross-organization capability bundle. No bilateral integration required. Trust comes from cryptographic verification, not from contracts about agent behavior.
Multi-Agent Pipelines
ACP packages: pkg/dcma · pkg/agent · pkg/ct · pkg/lia
A human operator authorizes a planning agent, which delegates to a research agent, which delegates to a data-retrieval tool. ACP's delegation chain model (DCMA) maintains a cryptographic link at every hop: human → planning-agent → research-agent → tool. If the tool causes downstream damage, liability attribution is unambiguous — the entire chain is verifiable and auditable. Any agent in the chain can only re-delegate capabilities it itself holds; it cannot escalate its own authority.