The Monday unlock checklist: fail-closed before your agents fan out

Monday is when agent fleets awaken eager for new tools and functionalities. However, the failure to follow a proper unlock checklist can lead to unexpected errors and surprises. Here is a fail-closed Monday unlock process that can help ensure a smooth rollout.

First, prove the latch by ensuring that the agent will only dispatch if certain conditions are met. Missing, unknown, or unvalidated fields, unrecognized permission verdicts, or unlisted tools should prevent the agent from executing. This is a strict approach, as opposed to best-effort or assuming the model will figure it out.

Before any rollout, make sure to validate tool definitions and structured output schemas, as these are stored in source control. If tools have been updated since Friday, assume that caches and clients may not have the latest changes. Validate the information before planning any actions. It's crucial to run a dry-run or DRY_RUN before enforcing any changes. This process should be followed by canary testing with one agent, then scaling to the entire fleet.

Next, separate who may call from what may steer after. Treat tool enablement as an allowlist rather than whatever the server advertised. Unknown permission verdicts should be denied, never coercing to auto-approve. Any destructive actions, such as writes, network access, or administrative tasks, should be denied until a human explicitly approves.

When enabling parallel or coordinator modes, make sure to set caps on retries and total tokens, as well as session budgets. This will prevent loops from occurring and dying politely. Also, pin model IDs to avoid using provider defaults, which are not a part of the release process.

For the actual Monday checklist, verify that the following items have been completed:

- Tool and output schemas have been diffed and validated (fail loud on drift)

- Dry-run and DRY_RUN have been green

- Tool allowlist has been reviewed, with unknown entries denied

- Destructive tools require explicit request; read-safe actions are automated

- Tool/web output is treated as data, with no privilege granted from retrieval

- Model IDs are pinned, with session token/retry caps set

- Parallel or coordinator spend is capped before fan-out

- One human-facing error path is implemented, avoiding eternal apology loops

If any item on the checklist is unchecked, do not unlock that surface. Ship less and stay fail-closed.

It's essential to remember that agents can be clever within the fence, but the fence itself should never be optional. By following this fail-closed Monday unlock checklist, agents can move fast while still being defensible during standup meetings.