Deterministic triggers for agent work

Research question

In a system where AI agents act for a small company, which decisions must stay deterministic — and what does the system look like when “is there work?” is answered without a model call?

Background

Agent frameworks often run a model in a loop and let it choose the next action, including whether to do anything at all. That makes cost and behaviour depend on the model’s judgement at the least observable point in the system. Acthelm takes the opposite position: no work means no model call.

Method

Separate the system into three layers and keep the model out of the first one:

1. Detection (deterministic). A cron-driven reconciler compares goals with measured metrics and emits tasks from a fixed rule set (GOAL_BEHIND, GOAL_UNMEASURED, GOAL_NO_WORK, GOAL_CONFIG_INVALID, GOAL_ACHIEVED). At most one open task per goal and rule, enforced by a partial unique index over open tasks (uq_tasks_open_dedupe).
2. Execution (model-driven, bounded). A worker claims a task with a conditional UPDATE … WHERE status = 'READY' and calls a model CLI in an isolated container.
3. Authorisation (deterministic + human). Signed policy, an allow-list of low-risk paths and a human signature decide what reaches production.

Implementation

See the Acthelm case study for the architecture. The relevant pieces are app/business_reconciler.py, app/worker.py and the migration that adds partial unique indexes for open work (migrations/versions/b8e41d0c7a52_open_work_unique_indexes.py).

Observations (so far)

• No model calls when idle. With no tasks, the worker sleeps; there is nothing for a model to be asked.
• Deduplication belongs in the database. A partial unique index over open tasks is a stronger guarantee than asking a model or a scheduler not to create duplicates.
• Crash recovery needs checkpoints at the model boundary. Storing the result of a completed model call means recovery never repeats it — important when calls are slow, quota-limited or not idempotent.
• Infrastructure failures are not task failures. Quota, rate-limit and metered-key refusals are provider health states with a cool-down, not task outcomes.

These are design observations from building and operating one system, not measured comparisons against an alternative design.

Limitations

• One system, one company, one host. No controlled comparison with a model-driven planner.
• Fixed rules only find the kinds of work someone anticipated.
• The unattended code-change path is not yet enabled, so the execution layer has mostly run attended.

Next steps

• Record, per rule, how often a created task led to an accepted change.
• Run an unattended pilot of the code-change stage with recorded evidence.

References

• Acthelm repository: github.com/Seerxo/Acthelm — see README.md, docs/AUTONOMOUS_AGENT_ARCHITECTURE.md, docs/operations/DECISIONS.md