Agent Harness in 7 Days: Let an Agent Run Unattended Through the Night

Why an Agent Starts Degrading Once It Runs Past the Second Context Window

Start with what really separates a long-horizon task from an interactive conversation, then reproduce the degradation with your own hands in a run that crosses two context windows: the second window redoes everything the first one already did. The model did not get dumber. There is simply no evidence left in its context that any of that work happened.

The Line Between State and Context: Moving the Authoritative State Out of the Window

Fix yesterday's degradation: write the authoritative progress to disk, and after the window resets rebuild an equivalent context out of what is on disk. The point is not how to write a file. It is where that dividing line belongs, and why the goal of rebuilding is equivalence rather than restoration.

The Initializer Agent: init.sh, a Progress File and the First Commit

The first minute of an unattended run decides how efficient the next several hours are. Today initialization becomes a phase of its own: a script that gets the environment running, a progress file written for whoever comes next, and a clean initial commit, so that any fresh window is productive within three minutes.

One Thing at a Time: The Feature List and git Discipline

Today opens on a reversal. The hand-built Coding Agent course ruled explicitly that the agent must never commit to git, while in this course a git commit is the authoritative record of progress. Both rulings are right, because the premises differ. Get that distinction and the rest follows: list-driven task selection, and rolling back to the last verified commit.

Self-Verification: The End-to-End Gate, Premature Completion Claims and Automatic Intervention on Doom Loops

The two most expensive unattended failures are the premature completion claim and the doom loop. Today the harness gets an end-to-end gate and stall detection, and the emphasis falls on what happens after detection: switching tasks, rolling back, and stopping with an alarm each have their own test.

Crash Resume, the Budget Circuit Breaker and Governance Decay

Overnight the power drops, the process gets killed, the money burns through, and the rules you set at the start quietly get forgotten. Today adds three things: resuming after a hard kill, halting the moment the budget is blown, and keeping the constraints alive after compaction. That last one is a real risk somebody has already measured.

Putting It Together: Let It Run All Night, Then Check the Work in the Morning

Merge six days of modules into a runtime you could hand to a colleague, run it long-horizon for real, then produce an overnight report a person can read in three minutes: what got finished, where it is stuck, why it stopped, and what a human has to decide next.