Blog
Field notes on keeping agents honest — state drift, deterministic verification, and gating the push.
The Agent Wasn't Wrong About Its Own Work. It Was Wrong About Yours.
Three trials of parallel agent sessions produced fourteen recorded incidents, and the pattern under most of them was the same: a session acted on a fact about shared state that was true when it looked and stale when it acted. The dev server made it worse by confirming the belief, because it reads the working tree and a push reads HEAD. The one-line check that separates them, on real output.
18 Aug 2026 · 9 minTwo Registers, One Wall: In-Flight State That Never Gates a Push
Two agent sessions on one repository share a directory tree but not a clock, and between session boundaries nothing records who is halfway through which path. A second in-flight register closes that gap. It is machine-local, self-ignored by git, and structurally unreachable from the push gate — because the moment drift can be asserted from ephemeral data, the gate stops being provable against git. Shown on real casp output.
22 Jul 2026 · 8 minLoop Engineering: The Boxes in the Loop That Can't Be the Model
Loop engineering breaks an agent loop into four parts — trigger, topology, verifier, stop rule. A model can serve as three of them. The verifier and the topology check are where a model grading its own work quietly fails, and both need a floor that resolves against git. Shown on real casp output for a drifted state and an unexecutable plan.
21 Jul 2026 · 9 minA Verification Rule Isn't Shipped Until Existing Repos Can Adopt It
A new drift rule reddens every repo that already uses the checker, for history written before the rule existed. The fix is an adoption window derived from the project's own append-only data, not a config flag — then the second-order trap, where the rule shipped a template no installed user could obtain, and the command written to deliver it safely shipped two ways to lose data.
21 Jul 2026 · 9 minThe Plan Is a Queue: Handing an AI Coding Agent One Session at a Time
Handed a whole spec, a coding agent picks a plausible starting point rather than the next one, and every session opens with the same re-briefing. An ordered queue of session prompts in the repository fixes that — and introduces a pointer that can name a slice already shipped, which is why the handoff has to refuse mechanically rather than politely.
15 Jul 2026 · 6 minThe Machine-Readable Handoff: A Drift Verdict an Agent Can Act On
The record a session leaves behind is read first by the next agent, not a person. So the check on that record should return structured output an agent can act on — expected vs actual per finding, as data — while keeping every property that makes it a deterministic floor: no model, no network, resolved against git.
13 Jul 2026 · 7 minSelf-Verifying AI Agents: The One Check the Harness Cannot Run on Itself
The self-verification loop closes every gap except one — nothing in it verifies the harness's own record of the project against git. Why a second review agent is a second opinion rather than a ground truth, shown on a fabricated session close that every code check waves through.
13 Jul 2026 · 7 minPre-Push Hooks for AI Coding Agents: Gate Both Boundaries of the Session
Advice-level verification does not survive autonomy. How casp gates both boundaries of an agent session — casp next refuses to start on drift, and a pre-push hook blocks the push — so a false recorded state can neither orient the next agent nor leave the machine.
11 Jul 2026 · 7 minState Drift: Why AI Coding Agents Lose Track of Your Project — and the Deterministic Fix
Across many agents, sessions and models, the state you recorded quietly diverges from what git contains. Why it happens, why storage-and-recall tools don't close it, and how a deterministic check validates state against git and blocks the push on drift.