Shackle Brings Deterministic Runtime Governance to AI Agents
Shackle introduces SP/1.0, an open, hash-verifiable conformance standard for real-time governance of AI agent tool calls and certification.
Shackle is a runtime governance layer that mediates every tool call made by an autonomous AI agent in real time, halting runaway token loops, unchecked tool cascades, and budget overruns before the next call fires. Each mediated action resolves to exactly one verdict — ALLOW, DENY, or HITL — with a deterministic, inspectable reason.
Underpinning this is SP/1.0, a hash-verifiable conformance standard built on a formal model (Valid(τ) ⇔ Required(τ) ⊆ Supported(τ)) and a core invariant stating that a recorded action is not proof it was ever authorized. Conformance is checked against 15 public test vectors and a stdlib-only reference implementation, so claims can be reproduced with pytest rather than taken on trust.
The project also defines three certification tiers (Core, HITL, Sovereign) backed by an open, reproducible registry, and ships integrations for LiteLLM-based stacks including CrewAI, AutoGen, and LangGraph. For engineers, it turns agent safety claims into something independently verifiable rather than an unfalsifiable promise.