BrainHarness Autoresearchv0.3.0

An engineer responsible for an internal customer-support skill is frustrated by vague, apology-heavy replies. They build an eval.json with regex checks for ticket IDs and LLM-judged actionability, then let the autoresearch loop mutate the SKILL.md prompt over 30 experiments until pass rate climbs.

A platform team maintains several prompt variants for a code-review assistant and needs objective evidence for which performs best. They define word-count and contains evals, run autoresearch with --runs 5 for statistical significance, and compare scores in results.tsv before shipping the winner.

A search skill occasionally hallucinates placeholder URLs and returns thin summaries. The team adds banned_phrases and word_count evals to catch these failure modes, then uses autoresearch to find a prompt that keeps good behavior while eliminating the hallucination pattern.

A media-monitoring skill must handle both English and Chinese requests with equal quality. Using mixed-language test_inputs and a contains eval for localized summary markers, the team runs auto-improvement to discover a prompt that triggers consistent structured summaries in either language.

A skill author preparing to publish on the BrainHarness marketplace wants measurable proof of prompt quality. They build a small eval suite covering depth, sourcing, and tone, run autoresearch to derive an optimized prompt, and attach the results.tsv evidence to the listing.

The skill and eval tooling stay free and MIT-licensed to drive adoption, while a hosted service executes long autoresearch sweeps on managed GPU/LLM infrastructure and stores versioned results. Teams pay for compute hours and result history rather than for the code.

The skill is packaged into an internal platform that connects to a company's private skill registry, enforces eval standards, and runs scheduled optimization jobs with audit logs. VPC deployment and SSO justify a per-seat or per-workspace license.

Individual developers get a limited number of free experiments per month in a web UI that generates eval.json and visualizes dashboards. Power users and small teams upgrade to unlimited experiments, private eval suites, and API access billed by consumed tokens or experiment count.