Tombstone
OpinionPanel.AI
Mar 2026 – Jul 2026 · age 44
What it was
A research assistant for health questions, built on Jeff's own Adama platform in seven weeks of spring 2026, and stated plainly to be not a medical advice tool. Each user got one private living document. A helper agent with 44 tools filled it from chat and from uploaded files: labs, medications, history. A question went to a panel of AI specialists, each a durable session on its own row, who debated by name across up to three rounds. A reviewer agent decided when to stop, and a letter writer produced a cited letter to bring to a doctor, exportable as PDF. It worked end to end on staging. The landing page never left "Coming Soon", and billing was never wired.
Wins, for the age
- At forty-four, Jeff made agents a construct of the language itself, with typed tools and a compiler that refuses an agent without a round cap. No popular equivalent has claimed that idea yet.
- Gave every specialist row its own agent session, so a panel of many was one type, finding on their own what Cloudflare's Agents SDK had already found.
- Ran the debate as a durable state machine that could crash or sleep and resume mid-round, the shape LangGraph's checkpointed graphs had already shown.
- Kept private data structurally out of reach: one document per user, and an admin plane that could not see inside.
What it taught
- The platform's board-game guarantees (hidden information, durable waits, crash recovery) fit a sensitive AI product with no new infrastructure.
- The reviewer is the hard agent. The stop decision needed tools, a fallback, a default, a round cap, and finally a cheaper model with no retries.
- Untrusted text must never reach an agent that can write, and code after a state transition still runs. The primate built that trap into their own language.
- Ship the legal posture with the code, and give hand-rolled security plumbing a second reader.
Genealogy
Ancestors: Adama: the platform it ran on. Adama Agents: durable agent sessions with typed tools. Adama Language: private-by-default records and state machines. Adama Rxhtml: the routes. Harvester: turning a site into template plus data. Tutor: a machine that explains its reasoning. Descendants: Hearth: the next fire, lit smaller and more honest.
Epitaph
Here lies OpinionPanel.AI. It ended the way it worked: seven invented voices debated its future for five rounds, recommended a pivot, and the author's last commit laid it to rest.