Skip to main content
Version: 1.0.27

Governor

The Governor is a senior reviewer that watches a long run and keeps it on plan. Think of it as a second, more experienced pair of eyes — one that runs on a more capable model than the agent it oversees: while the agent does the work, the Governor periodically checks that the work still serves your goal and stays grounded in the facts — and steps in only when it needs to. This page explains what the Governor does, how it appears in chat and workflow runs, and how you can steer it yourself.

What the Governor does​

Long, multi-step runs can drift. An agent can chase an interesting tangent, over-invest in one sub-question, or gradually wander away from what you actually asked. The Governor exists to catch that early.

At regular checkpoints during a run, the Governor reviews the agent's progress against two things:

  • Your goal — the objective you set for the conversation or workflow step.
  • The grounded facts — what the run has actually retrieved and computed so far, rather than what it assumed.

At each checkpoint the Governor issues one of five graded directives, from least to most forceful. Most of the time it's the first one and you never notice:

  • Proceed as you were — the run is on track. The Governor stays silent and invisible; nothing is injected and the agent carries on.
  • Proceed, mindful — the run is fine, with a gentle heads-up note. The agent keeps full autonomy.
  • Further clarification requested — the Governor asks the agent to explain why its next step is still on-plan, and a short back-and-forth follows; a good justification lets the agent continue.
  • Course-correct — the run has drifted. The Governor issues a firm, evidence-grounded correction telling the agent what to change.
  • Finalize now — the run is rabbit-holing without making progress. The Governor brings it to a close and has the agent deliver the best answer it can from what it already has, rather than burning time on a dead end.

Each intervention is graded to the severity of the problem, and the Governor has to justify its reasoning before it issues a directive — so a steer reflects a considered judgment about the run, not a reflex.

The Governor defaults to non-intervention

The Governor is deliberately conservative. On a healthy run it simply watches and lets the agent work; most checkpoints pass with no steer at all. Its job is to make long runs more reliable, not to second-guess every step — so when it does intervene, it's because the run genuinely needed it.

Because the Governor runs alongside the agent rather than blocking it, its checks don't stall your run — the agent keeps working, and a steer is applied on the next step.

How the Governor knows what "good" looks like​

The Governor doesn't judge a run on opinion — it judges it against a shared record of grounded facts drawn from your own AMPG Knowledge Graph.

When a run needs domain knowledge, the agent queries your AMPG Knowledge Graph — the Augmented Multi-Resolution Property Graph built from the documents you've added to AlphaAgent. Every fact retrieved that way is written to a shared, run-wide grounding record (GROUNDING.md): the cited facts, taken from your material, that the run has actually established. This record is the run's source of truth, and it traces straight back to the documents you provided — the grounding is never invented, only what your AMPG Knowledge Graph returned.

Here is the key idea. The specialist agent doing the work reads that grounding record — but it is also deep in the task itself: writing and running code, chasing sub-questions, carrying a large and ever-growing working context. That immersion is exactly what makes an agent lose the thread: the goal gets buried under the mechanics of the work.

The Governor reads the very same GROUNDING.md, but carries none of that working context. It sees the established facts and your goal, without the noise of the in-progress work. That deliberate asymmetry is what lets it stay high-level and catch drift the specialist can't — the worker is down in the weeds, the Governor is looking at the map. And because both are anchored to the same grounded facts, when the Governor steers it is citing the very evidence the agent is meant to be working from, not overruling it with something new.

In short: the grounding is yours. It is built on the documents in your AMPG Knowledge Graph, shared between the agent and its reviewer, and it is what makes the Governor's oversight trustworthy rather than arbitrary. (On a run that doesn't draw on an AMPG Knowledge Graph, there is simply less to ground against, and the Governor falls back to checking progress against your goal and the run's own trajectory.)

Where you see the Governor​

When the Governor is active on a run, a Governor panel appears in the run view — in chat for a Deep Traversal conversation and in a workflow run. The panel shows the Governor's activity as the run progresses:

  • Checkpoints — each point where the Governor reviewed the run, with its assessment of whether the work is on plan.
  • Steers — any nudges, course-corrections, or finalize decisions it issued, together with the reasoning behind them.

This makes the oversight transparent: you can see exactly where the Governor looked, what it concluded, and what (if anything) it told the agent to change.

Steering it yourself​

The Governor keeps humans in the loop. You don't have to wait for it to act — you can add your own guidance to a run at any time, and you stay in control of it:

  • Add a steer. Give the run your own instruction — refocus it, add a constraint, or tell it what matters most. Your steer is applied to the agent just as the Governor's own would be.
  • Edit a steer. Change the wording of a steer you added before it takes effect.
  • Withdraw a steer. Remove a steer you no longer want, so it never reaches the agent.

The "Apply now" grace window​

At a checkpoint, a steer isn't rushed onto the agent instantly. There's a short grace window during which a pending steer is held before it's applied — giving you a moment to review it, adjust it, or withdraw it first. If you want a steer to take effect right away, use Apply now to skip the wait. The grace window means the human always gets the last word at a checkpoint, without slowing the run down when you're happy to let it proceed.

Consumption​

The Governor is a reviewer running in parallel with your agent, so its checks use model capacity too. That usage is counted in the run's totals alongside the agent and the coder, so the token figures you see on a run reflect the full cost of the work — the agent, its analysis, and the oversight — with nothing hidden.

A deliberate — and economical — trade-off. The Governor runs on a more capable (and more expensive) model class than the worker — for example, a fast, cost-efficient Sonnet-class agent doing the heavy lifting, overseen by a stronger Opus-class Governor. The economics work precisely because the Governor is used sparingly: it weighs in only at checkpoints, and most checkpoints pass with no steer at all. So you get the sharper judgment of a top-tier model applied exactly where it pays off — spotting drift and keeping a long run on plan — without paying to run the whole task on that premium model. A small, metered amount of high-end oversight steers a large amount of cheaper execution, and it typically saves far more than it costs by cutting short the wasted, off-plan work that would otherwise burn tokens on a dead end. See Code interpreter and datasets.