Impact Boundary Labs
GitHub
Documentation navigation

Core docs

Limitations

The public v0.1.0 bundle is intentionally bounded. It demonstrates one idea clearly - agent writes pass through a boundary before they become real external effects - without pretending that a local runtime solves the larger production problems around agents.

What the v0.1.0 bundle can claim

  • It controls agentic write impact by separating intent from materialization.
  • It creates a deterministic boundary before external effects exist.
  • It checks state and policy before admitted work becomes claimable.
  • It demonstrates the boundary flow in a local runtime with Core inside the Adapter Host.
  • It makes external impact explicit, bounded, and inspectable: request, decision path, claimable work, outcome.

What the v0.1.0 bundle cannot claim

  • It does not guarantee safety.
  • It does not prove semantic correctness.
  • It does not replace human review.
  • It does not solve AI safety.
  • It does not fully verify agent behavior.
  • It does not provide a production-ready security layer, durable queue guarantees, hosted identity, or broad adapter trust.

Bounded is not safe

Bounded means the path is structured: the request is distinct from the decision, work is distinct from the external effect, and outcome is distinct from the original promise. Safe would imply much broader guarantees about hostile cases, operational resilience, and deployment context. The bundle claims the first, not the second.

Human review stays in the picture for the same reason. The boundary can route, constrain, and structure the path to impact - it does not remove the need for judgment in sensitive cases.

Beyond the local bundle

Moving from the local bundle toward production is separate engineering work with its own tracks: durable persistence, stronger identity and authorization, adapter trust and registration, review and escalation workflows, deployment and recovery, monitoring and audit, and operational security.

Those tracks matter, but they are not the boundary pattern itself. The right question for the current artifact is: does this model create a clearer, more controllable path from agent request to external effect?