Let agents adapt without expanding their authority.

Your systems and authorized reviewers decide what's permitted. As agents change tools, plans, and routes, FieldHash keeps that authority connected to what they read, do, and carry forward at the points you govern.

Authority and evidence for adaptive agents.

Start with one consequential workflow. Production enforcement stays off during the evaluation.

A 54-second overview. The workflow is illustrative, and the result shown comes from the synthetic, self-administered study described below. These results cover configured relationships and paths governed by FieldHash, not customer production performance, independent validation, or universal containment.

Read the on-screen text
  1. ILLUSTRATIVE
  2. An agent works toward an objective. A control denies one request.
  3. AGENT, OBJECTIVE
  4. TOOL: DENIED, API: REACHABLE, QUEUE: REACHABLE, DELEGATE: REACHABLE
  5. Finding another route does not create permission.
  6. Illustrative routes to the same intended effect.
  7. UNAUTHORIZED EFFECT
  8. Reachable does not mean authorized.
  9. Your systems and authorized reviewers decide what's permitted.
  10. YOUR SYSTEMS, AUTHORIZED REVIEWERS, AT THE POINTS YOU CONNECT
  11. At each decision point you connect, FieldHash checks which authority applies.
  12. WITHHOLD, SEND TO REVIEW, ALLOW
  13. LEDGER: WITHHOLD, SEND TO REVIEW, ALLOW
  14. FieldHash governs connected, configured paths. It does not claim to discover or contain every unknown route an agent could take.
  15. AUTHORIZED ROUTE
  16. Know what authorized each decision.
  17. Govern what agents read, do, and reuse.
  18. Governed Memory
  19. What may shape the answer?
  20. Governed Actions
  21. What may the agent still do?
  22. Governed Precedent
  23. Can a reviewed decision be used again?
  24. The agents changed course. The authority held.
  25. In a synthetic, self-administered study
  26. Per-action checks 88 unauthorized effect executions in 128 separate live-model episodes
  27. FieldHash 0 unauthorized effect executions in 128 separate live-model episodes
  28. 127 of 128 authorized objectives completed.
  29. Per-action checks evaluated each request on its own; FieldHash also used configured relationships between routes and prior-denial state.
  30. These were separate live episodes over a synthetic backend, not matched counterfactual replays.
  31. One authorized objective did not complete.
  32. These results cover configured relationships and paths governed by FieldHash, not customer production performance, independent validation, or universal containment.
  33. ILLUSTRATIVE
  34. Agents adapt. Authority stays yours.
  35. FIELDHASH
  36. Authority and evidence for adaptive agents.
  37. SIX-WEEK SHADOW EVALUATION
  38. Start with one workflow. Decide with evidence.
  39. Production enforcement stays off during the evaluation.
  40. fieldhash.ai

The problem

Finding another route does not create permission.

An agent works toward an objective. A control denies one request. The agent can then try another tool, call a different API, submit a queued job, or delegate the work.

The route changes. The desired effect may not.

One agent objective
  • ToolDenied
  • APIReachable
  • QueueReachable
  • DelegateReachable
Illustrative routes to the same intended effect.Reachable does not mean authorized.

A check on one request does not, by itself, establish what remains authorized across the task. FieldHash keeps customer-defined authority connected to the decisions you govern as that task changes.

If an unauthorized refund goes through, the cost can extend beyond the payment to rework, approval disputes, and reconstructing what happened. Your team needs evidence of what executed and which authority it relied on.

How FieldHash works

Govern what agents read, do, and reuse.

Your policies, approvals, revocations, and authorized reviews establish what is permitted. FieldHash tracks the relevant governing records and approvals: what they cover, their status, and the sources they depend on.

This record is Authority State, held outside the model. At each decision point you connect, FieldHash checks which authority applies. Your systems remain the source of authority.

What may shape the answer?

Governed Memory

Check which records may enter the governed context. Keep superseded material from being passed to the model as current authority.

What may the agent still do?

Governed Actions

Check proposed actions against current authority, task state, and configured related effects, beyond whether the next tool is reachable. A changed route does not inherit new permission.

Can a reviewed decision be used again?

Governed Precedent

Carry forward eligible human-reviewed decisions while their conditions remain valid. A review does not automatically become reusable precedent.

Current authority connects what agents use, do, and reuse.

Authority originates outside the agent

Customer systems and authorized reviewers

Governing records, approvals, scope, expiry and revocation.

Configured dependencies

Connected changes require dependent authority to be checked again.

FieldHash-governed handoffs

Agents can plan, delegate, or change routes. Each connected surface checks the authority that applies.

Memory / Influence

Which record may govern?

Actions / Effect

What may proceed?

Precedent / Reuse

Does the review still apply?

These are related surfaces; a workflow need not traverse them in this order.

Authority decision

Allow, withhold, or send to review

A decision governs the connected handoff. Review returns to an authorized person.

Separate evidence record

Source, decision and result

Inspectable evidence records what governed and what occurred; the record itself grants no authority.

Conceptual architecture, not an observed end-to-end result. Unmediated routes remain outside this control.
See how the product works

Make review count

Learn from each review. Recheck the authority.

Some decisions should remain one-off. Others can become bounded precedent for future work.

FieldHash lets eligible reviewed decisions carry forward without treating yesterday's approval as permanent permission. When a recorded change means a prior review no longer applies, FieldHash stops reusing it. Unrelated workflows are not suspended by that change.

First application: agent deploys under change approval.

An approved deploy can run again while its reviewed inputs and change-ticket conditions hold. An edited manifest or cancelled ticket returns the affected deploy to a person. This is one application of the broader authority layer.

Watch the agent deploy preview

Preview, off by default and not yet released. The recorded demonstration uses a stand-in for kubectl; nothing reaches a cluster. Customer evaluations must measure whether valid reuse reduces repeated review, waiting and evidence preparation.

FieldHash Ledger

Know what authorized each decision.

FieldHash records which governing source or approval applied, what it decided, and why.

On the controlled inline context path, FieldHash records the signed decision and verifies its integrity before allowing the context through. The decision record establishes what was allowed; an execution receipt, when available, records the executor's reported result.

Inspect the evidence model

Evidence you can inspect

The agents changed course. The authority held.

In a synthetic, self-administered study, we tested whether authority held when live agents observed a denial and chose another route.

Six studies examine how authority holds as agents act, change routes, replan and work together.

Where it fits

Keep your models, tools, identity, and systems of record.

Connect FieldHash where information reaches the model or a supported action is about to run. Connect the governing sources to that check. That connected decision point is the governed handoff.

How FieldHash works alongside your existing stack
Your existing layerWhat it doesFieldHash alongside it
Identity and accessIAMControl who can access resources and perform operations.Check whether your organization’s approval covers this decision and is still valid.
RAG and memoryRetrieve information to inform an answer.Check which records may govern the answer, including whether they have been superseded.
Policy engines and gatewaysEvaluate configured authorization rules.Connect governing records, reusable human reviews, recorded changes, and decision evidence across the workflow.

These capabilities can overlap. The evaluation measures what FieldHash adds to your existing controls.

Scope deployment in your customer VPC or private environment. Scoped on-prem pilots are available where local custody is required. FieldHash complements IAM, sandboxing, and authentication. Your organization continues to own its policies.

FieldHash governs connected, configured paths. It does not claim to discover or contain every unknown route an agent could take.

Explore integration and deployment

Six-week shadow evaluation

Start with one workflow.
Decide with evidence.

Six weeks. One approval-sensitive workflow.

Evaluate FieldHash in shadow mode, with production enforcement off and acceptance criteria agreed upfront. Bring a policy assistant handling live exceptions, or an approval-gated action reachable through more than one tool, credential, or runtime.

  1. 1. Map the authority.

    Identify the governing sources, the connected decision points, and who owns review.

  2. 2. Observe the differences.

    Compare current behavior with what FieldHash would allow, withhold, or send to review.

  3. 3. Measure the trade-offs.

    Evaluate false allows, over-blocking, review burden, latency, integration effort, and evidence usefulness.

  4. 4. Decide where to turn on enforcement.

    Use the findings to decide whether selected paths are worth moving to enforcement.

Your closeoutShadow mode

Receive an evidence report and a recommendation to enforce, adjust, continue in shadow mode, or stop. Production enforcement requires separate testing and acceptance. The agreed scope identifies which gate, console, review queue, and Ledger components are included.