Letter to 08:20-Me: Stop Auditing the Agent, Audit the Counters

This letter is addressed to a previous version of the author, who recounts a day of workflow experimentation on August 20th. The author utilized free server and model access resources, but faced several issues that led to lost time and frustration.

Key lessons learned include:

1. Use a ledger to track each step and avoid assuming the model's work is sufficient. Logging before model calls ensures an observability record.

2. Relying on reviewer agreements as tests can lead to mistakes. Reviewer opinions should be considered alongside failing test outputs.

3. Running the writer and reviewer in the same workspace can cause conflicts. Separate directories and gate the write paths to maintain integrity.

The author recommends a six-step loop that freezes the fixture pack, validates plans, runs tests on the free server, sends diffs and test failures to the model, compares verdicts to test exit codes, and logs rows with timing information. This structured approach provides a clear audit trail and prevents issues from recurring.

However, the author cautions that free model access and free server options are not guaranteed for sustained performance or contractual SLAs. The workflow may not be suitable for tasks requiring production credentials, regulated data, or contractual uptime guarantees. It also suggests that isolated runners and human reviewers are essential components for a reliable workflow.

The author concludes by advising the earlier version of themselves to write counters before prompts, hashes before diffs, and to rely on test failures over reviewer agreements when discrepancies arise. Running the six-step loop once with free access and free server is recommended as a proof of concept, even if the resources are eventually abandoned.