Deterministic Back-Pressure

Type checkers, linters, tests and structural lints are the deterministic signals that make looping agents converge; once a rule is objective, encode it as a check rather than prose.

Evidence: moderate

An agent running in a loop will keep producing output whether or not it is getting closer to the goal. Without some signal that pushes back on wrong output, a loop drifts, and instructions written as prose are interpreted loosely or ignored once context fills up.

Practitioners converge on deterministic checks as the main source of that signal. Geoffrey Huntley's Ralph loop, a shell loop that re-feeds a fixed prompt to a coding agent, is viable only because type systems, linters and tests reject bad work on every iteration; he calls it "deterministically bad in an undeterministic world". OpenAI's harness-engineering team enforces architecture mechanically: each domain has fixed layers checked by structural tests, custom linters check documentation freshness and cross-links, and lint error messages are written to inject remediation instructions into the agent's context. Their principle is to enforce invariants rather than implementations. Martin Fowler, quoting others, notes that agents act on every warning a human would let slide, and argues that once a rule is objective it should move from prose to a deterministic check. He adds that the verification loop, not generation speed, is the advantage.

In the vocabulary of guides and sensors, back-pressure is the computational sensor. It works best when failures are returned in a form the agent can act on, as described in failures returned as actionable feedback, and it underpins long-running patterns such as one unit of work per session and verify before done. Moving an objective rule out of the instruction file and into a check is the same move as fixing the environment rather than the output.

There is a limit. Passing checks does not prove the work is sound where the checks do not reach, and Dex Horthy's unattended "dark factory" degraded within months for exactly that reason; see unverifiable loops rot. OpenAI's contrasting stance keeps blocking merge gates minimal at high throughput and re-runs flaky checks rather than blocking.

Back-pressure is not only for coding loops. An agent that takes operational actions needs its own: a health probe, a diagnostic command's exit code or a smoke test that must pass after each action. Where an operational rule is objective (for example, never copy a database file while it is being written), it belongs in a check or a tool guard rather than in the agent's prompt.

Sources

  • Geoffrey Huntley, ghuntley.com (date not captured) — https://ghuntley.com/ralph/
  • Martin Fowler, martinfowler.com, 29 Apr 2026 — https://martinfowler.com/fragments/2026-04-29.html
  • Ryan Lopopolo, OpenAI, Feb 2026 — https://openai.com/index/harness-engineering/
  • Dex Horthy, via BigGo Finance news write-up, 15 Jul 2026 (secondary) — https://finance.biggo.com/news/15099f5634f5ab9a