Appendix — The study: the harnesses evaluated #

This appendix shows the work: the full list of harnesses that went through the study, with where they came from (upstream repository), the exact snapshot that was read (fork/commit/snapshot — the materialization of the method's cutoff date, ch. 01 §6) and the link to the complete evaluation of each one. The instrument used in every evaluation is the same: the HARNESS_EVAL.md template (and FRAMEWORK_EVAL.md for frameworks), applied through systematic code reading following the benchmark methodology (in Portuguese).

How to read this table #

• Origin: the public upstream repository.
• Version/snapshot: the version or snapshot that was read.
• Fork/commit (cutoff date): the picture frozen in the GHDaru/* fork — this is what guarantees reproducibility (anyone can read the same commit) and materializes the method's obsolescence mitigation. The forks are synchronized by the scripts/sync-forks.ps1 script.
• Evaluation: the complete document (metadata, per-dimension scores with code evidence, diagnosis and "what to steal").

The 16 evaluated #

Extension ext-1 (2026-07-31): the first Radar→corpus promotion #

The corpus grew from 16 to 18 through the very path the book itself institutionalized: the daily Radar (in Portuguese) found the candidates (2026-07-31 sweep), the editor approved the promotion, the repositories were forked for frozen reading, and the same instrument (HARNESS_EVAL.md) was applied — round ext-1, without touching the round 1/2 snapshots. Both pass the inclusion test of ch. 01 §4 (open source + general-purpose harness + adoption/representativeness): Grok Build for the opening of a complete commercial harness; Pi as a deliberately atypical case (Yin's replication logic called for a minimalist counterpoint, and the corpus was missing one).

Extension ext-2 (2026-08-02): the second promotion — and a new category #

The corpus grew from 18 to 20 through the same path: the Radar confirmed QM in a primary source (2026-08-02 sweep) and the critical reading of a vendor marketing article led to Kimi Code, verified at the source; the editor approved, the repositories were forked and the instrument applied — round ext-2. Kimi Code enters under the same criterion as Grok Build (the second model vendor opening a complete harness — the pattern became a trend, ch. 14). QM did not fit any existing archetype and inaugurates the "organizational agents" category: the unit of design is the organization (scopes, audience-based permissions, consent, auditing), and the agent loop is a swappable engine — indeed, Pi itself, evaluated in ext-1, is a dependency here (package.json). Yin's replication logic called for exactly this: a case that would test the limits of the taxonomy.

Extension ext-3 (2026-08-02): evaluated and not included — the inclusion test at work #

The method documents refusals too. Traycer (Traycer AI) was nominated by the editor, forked and evaluated with the full instrument (fork GHDaru/traycer, commit 65fc3d7, MIT) — and did not pass the inclusion test of ch. 01 §4: the open repository (~513k lines) contains clients, a CLI and a remarkable orchestration protocol, but none of the four harness pieces — the Host that runs loop, context, tools and control is a signed closed binary with a mandatory cloud (the repo's own AGENTS.md states the Host and backends are not there). The full evaluation (18/36) stays on record: it is the study's best-documented case of "open source" as a client-distribution strategy, and the central evidence of the new Appendix — The supply chain, for which the reading yielded the map of 18 orchestrated harnesses.

Extension ext-4 (2026-08-06): the corpus reaches 21 — and the first synthesis confronted #

Prime Agent (Prime Intellect) arrived through an editor's tip on announcement day and was the first candidate to threaten an Executive summary rather than add an addendum: the launch claims that context compaction "forces the model to work around its own scaffolding." Reading the code upheld the ch. 04 synthesis and recorded a caveat — compaction was not removed, it was subordinated to the agent (see ch. 04).

The round's structural finding is another: Prime Agent is built on Pi — the same four packages, a LICENSE with dual copyright (Mario Zechner + Prime Intellect), a README crediting pi-mono. It is the fifth Pi consumer recorded in the supply chain appendix, and it closes an argument from ch. 12: the lowest-scoring system in the corpus (Pi, 26/36 — which refuses half the dimensions by manifesto) was the base a frontier lab chose to build the study's most radical harness on.

Consolidated diagnosis #

The per-dimension results (scores 0–3, with evidence) and the comparative diagnosis live in the Harness Comparison — including the interactive heatmap. Beyond the scores, each individual evaluation brings: the harness's observed archetype, its strengths with file paths, and the "what to steal" section (patterns worth carrying over to other harnesses).

Method note (ch. 01 §6): selection followed replication logic (Yin) — representative and deliberately atypical cases; the unit of analysis is the source code; the scores follow the template's fixed grid (feature analysis, DESMET). The expiration scoreboard for the predictions is in the History (in Portuguese).

See also: reference implementations beyond the ones evaluated here are catalogued in Awesome Harness Engineering — Reference Implementations.