SKILL.state: replacing the agent's transcript with a state object, and cutting tokens 16x

Date ingested: 2026-09-19 Source: X home feed via @beamnxw (Google paper, reported second-hand) Raw: raw/twitter/feed/2026-09-19-morning-ranked.json

Evidence caveat, stated up front: this page is reconstructed from a social post describing a Google paper, not from the paper itself. The mechanism is coherent and the reported figure is specific, but the numbers are second-hand and should carry a discount until the paper is read directly.

TL;DR

Almost every agent runtime in production keeps an append-only transcript: each tool call and each observation is concatenated onto an ever-growing prompt. That has two costs, one obvious and one not. The obvious one is that prompt length, and therefore price and latency, grows without bound over a long run. The non-obvious one is context poisoning: an early wrong observation stays in the window forever and keeps influencing the model long after it has been superseded. SKILL.state replaces the transcript with an explicit mutable state object. The model sees only the skill specification, the latest observation, and a structured state it updates deliberately. Reported result: context-poisoning failures eliminated and cumulative token usage down 16x at 100 execution turns, with a prompt footprint that is constant rather than growing.

The change

flowchart LR
  subgraph OLD[Append-only transcript]
    T1[Turn 1 obs] --> T2[+ Turn 2 obs]
    T2 --> T3[+ Turn 3 obs]
    T3 --> TN[+ ... Turn 100<br/>prompt grows without bound<br/>stale errors persist]
  end
  subgraph NEW[SKILL.state]
    SPEC[Skill specification] --> P[Prompt<br/>constant size]
    OBS[Latest observation only] --> P
    ST[Structured state object] --> P
    P --> UPD[Model writes an<br/>explicit state update]
    UPD --> ST
  end
  classDef input fill:#dbeafe,stroke:#3b82f6,color:#1e3a8a
  classDef output fill:#d1fae5,stroke:#10b981,color:#065f46
  classDef warn fill:#fee2e2,stroke:#ef4444,color:#7f1d1d
  classDef aux fill:#e0e7ff,stroke:#6366f1,color:#312e81
  class SPEC,OBS input
  class P,ST,UPD output
  class TN warn
  class T1,T2,T3 aux

The design move is to make forgetting explicit. In an append-only log nothing is ever removed, so the model must learn to ignore stale content, and it does that imperfectly. With a state object, superseding a fact is a write, and the old value is gone. The token saving is a consequence of that design rather than the goal of it, which is the part worth holding onto: the 16x did not come from a compression algorithm, it came from deciding that history is not the same thing as state.

How this relates to prior wiki state

This is the third distinct mechanism in a month for the same problem, which crosses this wiki's threshold for naming a pattern. Agent memory has been tracking the general failure, and the specific instances now line up: SoL-Pi (09-18) halved token traffic in a self-improving research loop and showed the saving compounds because it doubles how many improvement iterations fit in one budget; AgentZip (09-19) compresses sibling agent sandboxes against their shared template and cuts sandbox-owned memory up to 8.7x, against 2.1x for stock Linux; and SKILL.state keeps a single agent's own prompt footprint constant across a hundred turns. Three attacks on agent state bloat at three different layers, the loop, the machine, and the prompt, all in a fortnight. None of them cites the others. The pattern is that agent context has become the dominant cost term in long-horizon runs, and the field noticed simultaneously.

It also puts a number on something the harness thread has asserted qualitatively. The agent-harness-engineering page has argued since 08-13 that harness design, not model choice, is the dominant variable in agent cost. The 09-18 empirical harness study made that concrete by showing three of its four findings were conditional on which model was running, meaning hand-tuned harnesses are misconfigured for the models they now serve. SKILL.state is the same claim with a clean mechanism attached: one architectural decision inside the harness, made without touching the model, is worth 16x on tokens.

Gaps

• Reconstructed from a social post. The 16x figure, the 100-turn horizon, and the context-poisoning claim all need checking against the paper.
• "Eliminated" is doing a lot of work. Context poisoning through a state object is harder but not impossible: a wrong value written into state persists exactly as stubbornly as a wrong observation in a log, and arguably more so because it now carries authority.
• No reported accuracy comparison. Constant-size prompts could easily cost task success on problems that genuinely need the full history, and the interesting number is the token saving at matched success rate.
• The skill specification has to be written. This is harness engineering work that moves effort from run time to design time rather than removing it.