agent-autonomy-kit

agent-autonomy-kit is aimed at one specific failure: an agent that idles until a human types something. The interesting part is treating autonomy as a designed loop — decide what warrants proactive action, what the bounds are, and how results are surfaced — rather than just 'run a cron job'. Pair it with a bounded budget and a notification-hygiene policy so proactive work doesn't turn into noise.

Use case: Moving an assistant from purely reactive to a bounded proactive loop.

# A queue and a heartbeat: the two files that make an agent proactive The core insight of this kit is that an agent stops waiting for prompts when there is always work visible to it. That requires exactly two durable artifacts: a task queue with explicit states, and a heartbeat that does work rather than checking status. ## Give the queue states, not just items The queue template uses ready, in progress, blocked and done. The states are doing two jobs: they let a scheduled run pick up work without re-deciding what to do, and they make blockers visible as a category rather than as an item that has not moved. ```text tasks/QUEUE.md Ready -> next actions, ordered In Progress -> at most one or two, with an owner Blocked -> item plus what it is waiting on Done -> moved here with a date ``` ## Make the heartbeat produce work A heartbeat that reports status costs a model call and changes nothing. A heartbeat that pulls the top ready item and advances it changes the system. Distinguish the two explicitly in the job description, because the default behaviour of a frequently scheduled agent drifts toward reporting. ## Bound the run Continuous operation needs a stopping condition. The practical one is a budget: stop when the ready queue is empty, when a step count is reached, or when the next action requires a human decision. An autonomous loop with no stop condition is a cost incident waiting to happen. ## Overnight work wants an isolated session Long autonomous runs should not inherit your interactive context; schedule them isolated so the run is reproducible and its transcript is readable on its own the next morning.