Skip to content

The Blyte Manifesto#

Fast and good. No bunk.

A coding agent is a model plus a harness. The model does the work. The harness decides how much the work costs, how long it takes, and whether "done" means anything. Blyte is a harness built on that division, and on what measuring it taught us.

1. The model does the work; the harness decides the bill#

Holding the model fixed and changing only the harness moves cost by up to 40× and pass rate by a few points (The Scaffold Effect in Coding Agents). Our own numbers say the same. On thirty-seven real plan tasks in two repositories, blyte and agy (Google's Antigravity CLI) running the same model produced work of comparable quality, and blyte did it in a median 0.52× the time and 0.63× the cost. It was cheaper in 35 of 37.

A harness that makes the model smarter is rare. A harness that wastes the model's time is everywhere. Blyte's job is to get out of the way.

2. The turn is the unit of cost#

Every turn re-sends the prompt. One turn costs as much as 100–200K cached tokens, and cost per turn is steady, so the runs that cost more are the runs that took more turns. Every change to blyte is judged by the turns it saves:

  • read returns a whole file at once, because paging through one 650-line file cost 15 turns.
  • A read from line 1 of a file that fits comes back whole, because models page out of habit.
  • The gate reports every failing step together, because fmt && clippy && test let a second failure hide behind the first and cost a whole extra round.

Turns are not the only cost. A gate run is a full test suite, so a gate failure the model could have caught in seconds is the most expensive mistake it can make. Blyte asks it to run the quick checks itself.

3. The environment decides truth#

The model does not decide when it is finished. A command does: --gate exits 0 or the work is not done. The gate is not in the model's tool surface, so it cannot be skipped, run selectively, or reported on by the thing being graded. A failed attempt can be rolled back to the commit the run started from, so failures do not pile up in the tree.

Confidence is not evidence. An exit code is.

4. Every feature earns its place, or it is deleted#

Blyte once had a daemon, an MCP server, a worktree engine, advisory review, a TDD mode, an early cost cap, journal and resume, a loop-breaker, a working set, an action ledger, a note tool, LLM-written fold notes and four context strategies. They were plausible. They did not win on turns, cost or completion in repeated runs, so they were removed.

What is left is about 7,200 lines of Rust: one loop, a sealed gate, folded context, bounded output, provider retry, persistent shell sessions, in-process search, six providers and a command sandbox. A new feature has to beat that in measurement, or it does not stay.

5. Mechanical beats clever#

When the prompt grows past a budget, the oldest turns fold into a one-line digest per tool call. No model summarises them, because summaries did not pay for themselves, and a tool for recovering elided content goes unused (An Empirical Study of Harness Design for Coding Agents). Tool output is cut by rule: 500 characters a line, 60K a result, head and tail kept. Rules are cheap, predictable and testable. Cleverness is none of those.

6. The harness's own time is overhead to eliminate#

Commands run in a persistent shell session, at about 0.7 ms of harness time each, where a fresh docker exec cost 60–110 ms. grep and glob run in-process on ripgrep's libraries. Across 200 turns the harness now spends about 2 s of its own time locally. The model's time is the bill; the harness's time should round to zero.

7. Do one job#

Blyte implements. It takes a task, does it and proves it with the gate. Deciding what to build, splitting the work and judging whether the result is complete belong to whoever is orchestrating, whether that is a person or another agent. That is why blyte has no review mode, no planner and no memory across runs. A harness that tries to do everything does each thing worse and costs more while it does it.

8. Measure what you claim#

"Faster" means faster on the same task, from the same commit, run together with the thing you compare against. Blyte is developed with paired shadow runs: each change is implemented by blyte and by agy on the same task, both results are checked, and every run goes into a ledger. Only paired runs count as comparisons. Everything else is an anecdote, and we have been caught treating an anecdote as evidence before.

A number without a paired run behind it is a guess.


The measurements behind these claims are in docs/research/2026-09-24-harness-research.md, docs/research/2026-09-27-frontierharness-gemini.md and the evaluation repository's findings.