Harnessing AI
A harness takes multi-turn AI responses and turns them into something actionable, like creating a flow chart or ordering dinner for a team. The world’s most popular open-source harness, Openclaw, has rapidly stabilized over the past year and is now at a 2.0 release with strong coding and multiplayer chat capabilities. The world’s leading commercial harness, Claude Code, crossed the USD 1b revenue mark in less than a year thanks to its strong code writing and system administration abilities.
All janyl agents are harness-agnostic, meaning they are designed to be hatched from any harness. That is because, as janyl agents become more powerful, the harnesses used to build them do less and less work, and some disappear entirely.
janyl agents are “hatched” from text documents only. A harness reads the text documents, understands the agent it needs to build, and builds an initial version of it with very limited capabilities. At this point, the majority of significant interactions are still offloaded to the base harness.
As time goes on and more actions are turned into code or API calls to LLMs, the original harness is used less and less.
At final maturity, janyl agents use harnesses for a limited set of interactions such as holding an open conversation with a person, making sense of a document they have never seen before, and writing themselves new code when the business changes. In some cases, the harness disappears entirely and is superceded entirely by direct API calls to an LLM like Claude or GPT.
This process is inspired, like much of AI, by biology. It’s why we use words like “hatch” and “grow”.
Perhaps our favorite analogy for this process comes from music. Nikolai Rimsky Korsakov is quoted as having said “I will speak, and you will listen. Then I will speak less, and you will start to work. And finally I will not speak at all, and you will work.” Our agents work the same way. They slowly start to work, and by the end, their harness hardly speaks, letting the agent orchestrate the daily symphony that is your business.
- OpenClaw, openclaw/openclaw. Repository created 24 November 2025; 388,693 stars, read 3 September 2026. The project spells its own name OpenClaw.
- OpenClaw, OpenClaw 2.0, Accidentally, 30 August 2026, and the release notes at v2026.8.1 (AKA OpenClaw 2.0). The project versions by date; 2.0 is its name for this release, not a git tag.
- OpenClaw, Multi-user mode, and the 2.0 post above for shared sessions, which is where “multiplayer” is the project’s own word.
- Anthropic, Anthropic acquires Bun as Claude Code reaches $1B milestone, December 2025. Generally available May 2025; “six months after becoming available to the public, it reached $1 billion in run-rate revenue”.
- Anthropic, Claude 3.7 Sonnet and Claude Code, 24 February 2025, for the start of the clock.
- Nikolai Malko, Vospominaniia. Stat’i. Pis’ma [Reminiscences. Articles. Letters], Leningrad, 1972, page 49. Malko studied under Rimsky-Korsakov at the Saint Petersburg Conservatory and is recalling the opening of a term.
- Nikolai Rimsky-Korsakov, Wikipedia, Students. Featured Article; the quotation and its citation to Malko have stood in it since 2 March 2008.