CORE: Open-source governance layer blocks AI agents from dangerous code changes

Key takeaway

• CORE is an open-source governance runtime that constrains AI agents with machine-enforced constitutional rules, blocking invalid code mutations automatically before execution.

• The tool makes autonomous AI workflows auditable and deterministic by logging every action's decision chain and enforcing architectural invariants structurally, rather than detecting violations after the fact.

• It is available now as a pip-installable package.

3 Key Points

FAQ

How does CORE actually stop an AI agent from making a dangerous change?

CORE runs the AI agent inside its governance system, never above it. Before execution, generated code is validated against constitutional rules (215 rules across 15 engines). If a violation is detected, execution halts with no partial state change and logs the violation. The agent cannot proceed until the violation is remediated or approved by the proper authority.

What kind of rules can CORE enforce?

Constitutional rules block unconditionally. Policy rules block when strict_mode is enabled, otherwise report. Advisory rules and capability tier currently report violations without blocking (blocking for capability tier is under review in ADR-079). The governance directory itself (.intent/) is immutable—no component can rewrite the rules.

Can I see an audit trail of what CORE approved or blocked?

Yes. CORE records every action with a complete consequence chain showing finding, proposal, approval authority, execution time, and file changes with commit hashes. Two examples are shown live in the documentation: one for autonomous self-approved changes and one for changes requiring human governor approval. Both are queryable end-to-end.