Without a governed boundary
An agent may reach tools, credentials, and services through paths that bypass a policy check. A policy recommendation alone cannot block those paths.
What changes when an agent acts?
With Anthesis
The agent proposes an exact action. Anthesis evaluates its policy, identity, scope, and approval requirements. The integration allows only the authorized effect and records its outcome.
Enforcement matters: Anthesis defines and evaluates the decision; the surrounding tool, gateway, credential boundary, or runtime must make bypassing it impossible for the scoped effect.
Try one governed action
Run a public reference trial that changes a file in a disposable Git repository, then attempts a raw-tool bypass and an out-of-scope write.
Run the reference trial
Clone Anthesis Governance Lab, then run from the repository root:
bash scripts/acquire-anthesis-lab.sh
bash scripts/run-reference-trial.shThe authorized write leaves a Git diff and decision/outcome evidence. Both bypass attempts are hard-denied with repository state unchanged.
What the result means
Runnable now The public evaluator and constrained reference harness exercise a real repository write and blocked bypass attempts.
Reference integration Dubnium is a personal self-hosted AI system demonstrating one bounded governed-agent execution composition; it is not required to use Anthesis or currently distributed as a general-purpose product. A public release may follow later.
In development Broader production enforcement profiles and live inference-integrity verification.
Limit: A local trial does not prove non-bypassability in an arbitrary runtime with raw shell, filesystem, network, or credential access.
Design-partner evaluation
What success looks like
Choose one low-risk, consequential action in your existing agent workflow and test it within an explicitly defined integration boundary.
Product goal: Help resource owners retain control over agent access, modification, and data transmission through explicit, enforceable authorization without replacing the model or agent system. Data sovereignty also depends on the surrounding runtime, network, storage, and legal or organizational controls.
- Authorization: A permitted effect executes, while denied and out-of-scope variants do not.
- Bypass resistance: Test alternative raw tools, credentials, and direct service paths for the selected effect; document remaining paths.
- Attribution: Connect the actor, requested effect, policy decision, execution, and observed outcome.
- Human approval: When required, the effect remains blocked until an approval bound to its exact scope is verified.
- Adoption: Integrate governance for that effect into the existing workflow and document configuration effort and tradeoffs.
Evaluation request: Run the public reference trial first. If the pattern is relevant to your system, identify one suitable workflow, its effect and bypass paths, and the evidence you need. Then assess it against the full trial criteria. The public reference trial does not itself verify your environment.
Integrate with an existing agent system
Anthesis is not an agent framework or an LLM runtime. It supplies governance decisions and evidence semantics; the integration controls which effect paths an agent can access.
Possible enforcement points include a constrained tool registry, MCP mediation, a gateway, downstream capability validation, or a sandboxed runtime. An SDK call without control of alternate paths may be advisory rather than enforcing.
Evidence, maturity, and limitations
Governance Lab independently tests deterministic decision contracts and recorded evidence. Its canonical, general, and inference-integrity scenario catalogs are separate proof surfaces; synthetic declarations do not execute production effects.
Reproduce the public proof surfaces →
Anthesis does not guarantee deterministic model output or universal production enforcement. The authorization decision must be enforced in the actual runtime. Evaluate a deployment against the published trial criteria.