Skip to content

B.3 Operating a GEE

A GEE has to turn an engineering priority into bounded work, preserve the evidence needed to judge the result, and recover when the work stops short. DocAble’s November 2026 workflow provides one concrete arrangement. Its strategy, epic, design, and dispatch documents have different consumers; they can be combined in another organization without losing the distinctions. The important connections are from priority to authorized scope, from scope to obligations, and from the resulting work to acceptance evidence.

In DocAble, I record requirements for the next deployment alongside longer-term priorities, and the orchestrator maintains the strategy document from my instructions. An epic develops a selected effort into scope, rationale, risks, relevant models, and a Definition of Done. Its machine-parseable phase table exposes dependencies and required reasoning capability to the scheduler. If review changes those dependencies, the schedulable plan must change too. Selecting a priority authorizes planning; it does not establish that the proposed decomposition is sound.

The design develops the engineering argument. It identifies the motivating failure or requirement, considers alternatives, and explains the proposed components and interactions. Stable invariant identifiers connect obligations to tests or analyses, while uncovered obligations remain explicit. A dispatch brief then assigns one ready phase, with a role, scope, time budget, relevant source, verification work, and expected hand-back. DocAble’s brief linter can reject missing structure and selected dispatch errors. It cannot decide whether the engineering premise is right.

From a Brief to Accepted Work

As an illustration, consider an assignment to improve PDF heading detection. Prepare the brief from the current repository, not from a remembered account of the component. Four kinds of information make it usable:

  • Intended change and preservation. Identify the heading behavior to improve, the existing behavior that must remain, and exclusions such as unrelated document mutations. Leave implementation choices open where they do not affect these obligations.
  • Engineering context. Point to current source, the representation the component operates through, and relevant invariants. Require the agent to query the models and checks associated with its refined target files before changing them.
  • Acceptance evidence. Specify evidence for the intended behavior, preservation obligations, and affected integration paths. Evaluate those properties independently of the implementation the agent happens to choose. Record unavailable evidence as an open question.
  • Budget and hand-back. State when to stop and what to preserve: commits, completed checks, partial results, unresolved questions, and unmet obligations. An honest partial result supports a decision about further work; an unsupported completion claim hides that decision.

The agent records its refined plan in the first commit and progress in later commit metadata. A sentinel check verifies selected prerequisites early: the expected worktree, branch, working directory, and registration marker. These checks can discover an unusable working environment before substantial effort is spent. They establish those prerequisites at that boundary; later integration still needs its own evidence. The companion epic, design-document, and dispatch-brief starters supply field structures that can be adapted to these consumers.

Completion review follows the work back into the repository. Check that the promised change exists, was integrated, and has evidence supporting closure. A complete sequence of commit markers helps locate unfinished work but cannot establish semantic correctness. If the work is incomplete, preserve what the next agent or engineer needs to resume it. Reconcile partial commits and current ownership before launching replacement work; otherwise recovery can create a second competing attempt while the first remains stranded.

Keep the Checking Process Operational

The factory’s own controls need the same scrutiny. In November, I retained cheap synchronous checks at merge and moved expensive validation into a periodic sweep. An intermediate design launched that sweep after every merge, so bursts of completed work still caused bursts of expensive validation and overloaded the host. Background execution had changed who waited without removing the coupling that caused the overload. The replacement gave the sweep its own schedule and lock, inspected the main branch in a disposable worktree, and allowed one sweep at a time. Broader validation remained part of release.

A configured schedule is not evidence that a check runs. The November 3 investigation found the timer in the configuration manifest but not installed. Checks had been running through a debounced post-merge hook, masking the missing schedule. Installation and inspection of the live schedule corrected that gap. Another failure concerned closure: an alert was marked resolved while its lint still failed. Rechecking before closure, or reopening the alert when the next sweep finds the condition again, was identified as necessary follow-up work.

Distinguish detection, acknowledgment, and repair in the operating record. The observed sweep detected a blocking failure and emitted an alert. A different alert, concerning accumulated worktree branches, caused a hook to refuse an orchestrator operation until acknowledgment. Those observations did not demonstrate detection, blocked dispatch, and verified repair for the same failure. Acknowledgment permits a deliberate response to a known problem; it does not make the problem disappear. Record the revision each check examined so that evidence for an earlier snapshot cannot silently validate later changes.

Make Recovery Improve Later Work

After recovery, ask what subsequent work should inherit. Was the problem missing knowledge, a stale representation, an implicit obligation, late evidence, or a mechanism that failed to act? The answer determines whether to improve a model, move a check, restrict an action, change the procedure, or retain deliberate human judgment. Record the obligation and intended consumer of the repair, then verify that the consumer actually uses it. This is the practical handoff from incident response to governance conversion.

The duplicate-dispatch incident in Appendix A illustrates this handoff. Its repair placed a completed-phase gate in work-order composition, so later assignments inherit the refusal automatically. The gate acts on recorded completion rather than inferring it from arbitrary repository changes; that boundary defines which recurrence the repair prevents.

Price that inheritance. A rare, inexpensive failure may not justify permanent machinery; a repeated reconstruction or recurring loss may justify it readily. Remove the superseded procedure or control when its replacement makes it unnecessary. Otherwise every repair adds another thing that later work must understand. The next run should inherit a better starting point, with the evidence needed to tell whether the repair worked.

© James C. Davis, 2026–present