Field Notes / Engineering

The Deterministic Spine: more reliable and more expensive

Field Note 001 · September 19, 2026

In The Deterministic Spine, I described moving repeatable work in my agentic workflow into scripts: setup, validation, commits, pull requests, cleanup, and merge checks. I expected fewer failures and lower token costs.

Before, the instructions described steps for the model to carry out. Now, they tell it to execute a script and use the result. The steps live in code, but the model still decides when to call it.

The September 17 comparison showed higher completion—and higher costs.

Figure 1. Before-and-after comparison, measured September 17. Repository mix differs; the later cohort is mostly workflow infrastructure work, so this does not isolate the spine’s effect.

We missed an ordering dependency: cleanup removed the worktree before the merge gate checked it. Twenty dispatches repeated setup and a full validation run to recover. We had made the scripts work against each other, and agents spent time and tokens recovering. That ordering was ours to fix.

The logs also showed duplicated mechanical work. Across 83 post-spine dispatch logs, agents spent 3,224 of 11,710 Bash calls on mechanical work the scripts were meant to replace. Reviewers repeatedly reconstructed diffs. Agents rechecked branch state and rediscovered toolchain settings.

We told agents to call the scripts but kept the old step-by-step instructions as fallbacks. Those instructions added context and invited agents to repeat the work themselves. Each fresh implementation agent read 3,725 words of legacy fallback procedures.

Next steps

• Move cleanup after the merge check so agents don’t have to recreate the worktree and repeat validation.
• Remove the old fallback instructions.
• Use shared commands to prepare diffs and commit fixes.
• Add help support and return the state and failure details agents currently reconstruct.
• Measure duplicated calls and compare completion and cost after these changes.