Governed execution for agentic systems

Let agents act without giving them invisible authority.

Anthesis evaluates consequential agent actions against explicit policy, authority, approvals, capabilities, and evidence requirements at the boundary where a host system invokes it.

The decision becomes meaningful when a tool surface, gateway, credential boundary, downstream validator, or runtime prevents the agent from bypassing it.

Two proof paths

Run the contract. Inspect the execution.

The evaluation and execution paths are deliberately separate so each can make a precise, testable claim.

Runnable now

Governance Lab

Run the immutable public Rust evaluator against deterministic declarations and inspect stable decisions, reasons, rules, and aggregate reports.

  • Reconciled themed demonstration packs spanning representative governed actions
  • Allow, approval-required, policy-deny, engine-guard, and expectation-drift outcomes
  • No GPU, live model, private credentials, or hosted service required
  • Declared effects are evaluated, not executed

Open the stakeholder walkthrough →

Reference integration

Dubnium governed-agent demo

Inspect a bounded tool-execution harness that normalizes actions, invokes Anthesis, resolves exact approval when required, executes constrained tools, and verifies evidence.

  1. Bounded action request
  2. Anthesis decision
  3. Exact approval when required
  4. Constrained Dubnium execution
  5. Sanitized, tamper-evident evidence

Open the governed-agent runbook →

How the components fit

Policy, validation, and execution have distinct owners.

Keeping these responsibilities separate prevents a demonstration harness or runtime from becoming a self-certifying policy authority.

Component responsibilities
Component Responsibility
Anthesis Policy authority, deterministic decisions, approval requirements, capability constraints, evidence semantics, provenance, and public evaluator contracts.
Governance Lab Independent deterministic conformance, scenario packs, reports, and stakeholder walkthroughs. It is not in the runtime critical path.
Dubnium Reproducible runtime, gateway, bounded tools, execution, and runtime evidence. It consumes decisions; it does not define policy authority.

Integration assurance

Where enforcement lives is not the same as how strong it is.

The same integration mode can be advisory, moderate, strong, or runtime-enforced depending on which bypass paths remain.

Anthesis integration enforcement locations and assurance conditions
Enforcement location Common modes Required bypass control Typical assurance
Application SDK wrapper Direct clients and raw credentials cannot remain an uncontrolled path. Advisory to moderate
Tool surface Tool wrapper / invoke, MCP mediation Raw tools, downstream MCP servers, and service credentials are unavailable or constrained. Moderate to strong
Infrastructure Gateway / sidecar, capability tokens Downstream effects are unreachable or reject missing, altered, expired, or out-of-scope grants. Strong
Runtime Sandboxed runtime Filesystem, network, process, credentials, and tools are unavailable outside governed paths. Runtime-enforced when complete

The governing question: What prevents the agent from producing this effect without crossing Anthesis?

Read the six detailed integration modes →

Current status

Use precise maturity labels.

Public materials distinguish what can be run today, what is a reference integration, and what remains under active development.

Runnable now

Deterministic public validation

Public evaluator acquisition, Governance Lab scenario packs, reconciled catalogs, aggregate reports, and stakeholder walkthroughs.

Reference integration

Bounded governed-agent execution

Dubnium demonstrates exact authorization binding, approval-gated constrained tools, source immutability, evidence sanitization, and manifest verification.

In development

Broader production assurance

Additional enforcement profiles and stronger live inference-integrity evidence, replay, verification, containment, and recovery.

Trust boundary: Governance Lab demonstrates deterministic contract behavior, not production effect enforcement. The Dubnium demo is a bounded reference integration, not proof that every Anthesis deployment is non-bypassable. Inference-integrity fixtures are not complete live exfiltration prevention.

Learn more

Choose the depth that matches your question.

Run Governance Lab

Execute and inspect the deterministic public proof surface.

Stakeholder walkthrough →

Inspect bounded execution

Follow the Dubnium governed-agent decision, approval, execution, and evidence path.

Governed-agent runbook →

Compare integration modes

Review mechanism, threat boundary, bypass assumptions, recommendations, and tradeoffs.

Integration documentation →

Plan an evaluation

Use bounded trial criteria and evidence expectations rather than broad product claims.

Trial criteria →

Read the project brief

A concise, printable explanation of the product, proof paths, and integration model.

Open the one-pager →

Read the whitepaper

The curated technical narrative, trust model, compositions, and adoption path.

Open the whitepaper →

Releases and source

Inspect immutable public evaluator releases and the canonical Anthesis implementation.

Public releases →
Anthesis source →