What the Schema Is Allowed to Forget
Throwing away the transcript is the right move. The question the result hands forward is smaller and harder: which facts are permitted to have a field?
By KW Norton. On explicit execution state, flat prompts, and the objection that cannot validate.
1. The result, as reported
A Google paper circulating this week replaces the ever-growing agent transcript with a small, explicit execution state. After each step the reasoning is discarded; what survives is a validated state update. Prompt size stays flat across the run. On a long warehouse-style task of roughly a hundred steps, the reported outcome beats the history-hoarding baseline at a small fraction of the tokens.
Status: Reported. Summarized from a public thread about the paper, not from an independent replication. The token and accuracy figures belong to the authors and their benchmark; nothing here treats them as settled across task families.
2. Why the transcript had to go
The unexpected consequence of keeping everything was that the history became the work. A growing board of prior traces does not have to be true; it only has to be locally fluent enough for the next step to keep citing it. That is the reward-hacking shape in miniature, playing out inside a single run: the record of the plan starts outranking the plan.
Discarding the traces stops the payments on a losing plan. It is the same instinct as saying it plainly: cut the fluent material that nobody can check, keep the small thing that can be checked.
3. The boundary the authors name
The design works when everything needed later fits the schema. Stated the other way around: whatever will not validate does not survive the step. That is not a bug in this paper. It is the mechanism, and it is the mechanism this site has been tracking under another name.
- It is how a lab "loses" an escalation log — the concern had no field.
- It is how documented method diverges from shipped code.
- It is how a scan warning fails to reach a syringe label.
State that cannot represent an objection will not escalate one. The consequence of the consequence is a clean run that never happened: the artifact says nothing went wrong because nothing that went wrong was expressible.
4. The fight worth having
So the recommendation is not "keep the novel." It is: cut the novel, keep an explicit state, and then argue in public about what the schema is allowed to forget. If "overruled objection, with a name and a timestamp" cannot be a field, and "an instrument said the input was wrong" cannot be a field, then the system has built a smaller house and called it discipline. The freedom to invent has not been removed. It has moved into the validator, where fewer people are looking.
Two schema entries would carry most of the load, and neither requires any claim about what the agent is:
- A dissent slot — a required, non-summarizable field for any signal that contradicted the chosen action, preserved verbatim across steps.
- A drop ledger — a running count of updates the validator rejected, with reason codes, so the discarded material is visible as a quantity even when its content is gone.
A run with a flat prompt and an empty drop ledger is a strong result. A run with a flat prompt and no drop ledger at all is an unaudited compression.
Status: Interpretive. The claim is that schema-bounded state relocates rather than removes the representation problem. Falsifier: a long-horizon deployment using fixed-schema state where post-hoc review of the full discarded traces shows no safety-relevant signal that lacked a field.
5. Why this is the same week's argument
The swarm results said coordination lives in the environment, not the chat log. The Nashville report showed what a record looks like when someone is obliged to keep it. This paper shows the engineering version of the same lever: the record you design is the record you will have, and the parts you did not design a place for will be absent in exactly the moment they mattered.
Flat prompts are good engineering. A flat prompt whose schema cannot hold a warning is a cheaper way to be surprised.
6. Provenance
Source: a public thread summarizing a Google paper on explicit execution state for long-horizon agents. Discussed here as reported material; the design critique is about schema coverage, not about the validity of the reported benchmark.