Trust & Governance

Autonomy you can audit.

AI agents that ship real work — behind approval gates, budget limits, and a field-level audit trail. Supervised autonomy, not a zero-human black box.

21748 sessions captured · 6148 memories indexed · every action attributable

See the control plane Enterprise overview

Two ways to run agents

"Zero-human" is a liability, not a feature.

Removing people from the loop doesn't remove accountability — it just removes your ability to answer for what happened.

Zero-human autonomy

  • ✕ Agents merge their own code — nobody signed off, nobody can say why
  • ✕ Destructive operations run the moment the model decides to run them
  • ✕ Spend is discovered on the invoice, not enforced up front
  • ✕ No memory of what the agent knew when it acted — post-incident review is guesswork
  • ✕ Compliance answer: "trust the model"

Supervised autonomy — FlukeBase

  • ✓ High-risk operations pause at an approval gate until a recorded decision exists
  • ✓ Autonomy bands — agents earn wider scope through demonstrated track record, never by default
  • ✓ Budgets enforced before the spend — preflight checks, per-project limits, a kill switch
  • ✓ Bitemporal memory — replay exactly what the agent knew at decision time
  • ✓ Compliance answer: the audit export

The control plane

Six governance layers. All built in.

Not bolted-on dashboards — the same engine that runs the agents enforces the controls.

Approval engine gate

Destructive and high-risk operations cannot run without a recorded decision. Rules route by risk; approvals are batched, attributed, and permanent.

Field-level audit trail

Every change to every entity is logged at field granularity — who, what, when, previous value. Exportable, with integrity verification.

Budget enforcement spend

Per-project and per-company budgets with preflight checks — the agent is stopped before the overspend, not reported after it. Kill switch included.

Autonomy bands scope

Agents start narrow and earn wider autonomy through a promotion history you can inspect. Every band change is a governed, logged event.

Supervised PR gate code

Agent-authored code passes a quality gate and lands through reviewed pull requests — CI, diff coverage, and human sign-off where policy requires it.

Auditable memory memory

Bitemporal, versioned agent memory. Ask "what did the agent know on March 3rd?" and get the actual answer — versions linked, nothing silently overwritten.

How a governed action lands

Propose. Gate. Execute. Prove.

  1. 1

    The agent proposes

    Deploy, payment, schema change, mass email — the intent is captured as a structured request, not a side effect.

  2. 2

    Policy decides who decides

    Inside the agent's autonomy band and budget? It proceeds. Above it? The request waits for a named human — on web or mobile.

  3. 3

    Execution is attributed

    The action runs with the agent's identity, the approval reference, and the budget line it drew from.

  4. 4

    The proof outlives the session

    Audit rows, change history, and the agent's memory state at decision time — available to your auditor, not just your engineers.

See the enterprise platform

Already on the platform? Open Control Center →