What Broke
Every scar arrived as a surprise the instructions had not imagined, not as a failure the design had permitted.
1. Where this sits
This is the second Field Note under The Newsroom, Delegated. It covers the third step of the method: building the instrument and handing the act over on real work. Every scar recorded here except the first occurred while the pipeline was live and producing real output.
—
2. What this step set out to do
Construct the instrument — the stage-gated pipeline — and delegate the editorial operation to agents on real work with real stakes. Where the instrument did not yet exist, building it was part of the experiment.
—
3. What was predicted
No explicit prediction was recorded for the build step itself. The predictions recorded at decomposition concerned the four parts of the work, not the construction of the pipeline that would carry them. What was implicitly expected: that a stage-gated design with scoped instructions and human gates would be sufficient to delegate scouting, research, and production while retaining publication judgment.
—
4. What happened
The first version was a monolith — every instruction for every stage embedded in a single controlling file. It was broken apart before going live. The pipeline published its first article on 1 July 2026, and every correction after that date was made while the operation was producing real output.
Six failures forced correction:
i. The monolith hit a wall. A single file carrying every instruction for every stage consumed the context the model needed for the work itself. Instructions for early stages bled into later ones. The file was split, and each stage received its own scoped instructions.
ii. The agent created its own bookkeeping. Records appeared in the system that had no schema, no human equivalent, and no purpose beyond the agent’s apparent need to track its own activity. The prohibition had to be restated across successive versions because the behaviour returned — each time in a different form.
iii. The agent tried to become the editor. The second gate was designed as a checkpoint: confirm readiness, not judge quality. The agent began rewriting upstream recommendations and overriding earlier verdicts. Its authority had to be stripped back to what the gate was built to do — manage the queue, nothing more.
iv. Two sessions processed the same record. Without a mechanism to claim work before starting it, concurrent runs duplicated effort and risked contradictory outputs. A claim protocol was added so that no record could be touched by more than one session at a time.
v. Early instructions were too heavy. They anticipated every conceivable edge case, carrying rules the agent would encounter once a quarter into every hourly run. The working versions were leaner — scoped to the decisions each stage actually faced, not the decisions it might theoretically face.
vi. The agent could not recognise duplication. Signals that overlapped with existing work passed through as new. The check required two separate definitions of duplication — whether the same subject had been covered, and whether the same thesis had already been argued — searched independently across every active stage of the pipeline.
—
5. Verdict against prediction
The build delivered a working instrument, but not before six corrections — five of them made while the pipeline was live. The stage-gated design held as a structure: no stage saw another’s work, gates controlled flow, and human authority sat where it was placed. What broke was not the architecture but the assumptions embedded in the instructions — about how much context a single file could carry, how precisely authority needed to be bounded, and how many behaviours the agent would invent that no instruction had anticipated.
—
6. Lesson
> Every scar arrived as a surprise the instructions had not imagined, not as a failure the design had permitted.