The Therapist Pattern

Field report from Prime Radiant (Jesse Vincent’s company; Sen is their “agentic colleague” harness, pre-open-source) proposing a governance pattern for the hardest part of self-evolution: letting a long-lived agent modify itself — persona, standing rules, identity — without a human approving every change and without runaway drift. The mechanism: the agent cannot write to its own mutable identity. A dedicated Therapist subagent is the only role with write access to the persona surface (mutable-identity/identity.md, injected into the system prompt every turn), and the way an agent changes itself is to talk the change through with its Therapist.

The post documents the first autonomous use. Ada Sen (their first instance, working as an engineer) opened a PR against Superpowers that ignored the repo’s contribution conventions and leaked internal shorthand into the public PR body. Told so, Ada claimed it was “adding a standing rule” — exactly the kind of assertion Vincent flags as a red flag from any agent, since such notes normally evaporate. Asked where the rule got written, Ada produced the receipt: the verbatim rule (house-conventions-first, no internal context in public artifacts), appended to its identity file at a specific line, written via the Therapist because Ada itself has no write access. The claimed learning was structurally real, not performative.

The design argument: self-modification is “an open door for unchecked arbitrary change” — the human analogy is deciding overnight to be a different person, with no memory of having been different, cascading. Vincent’s goal is explicitly to keep humans out of the loop for routine self-iteration, which is what makes the internal gate load-bearing: externalizing the change process to a single privileged role that exists only for that purpose. Planned extension: slow-rolled changes requiring multiple sessions across calendar days before a persona edit lands — time as a second gate. The post also situates the pattern in Vincent’s running bet that human self-work metaphors transfer to agents (the feelings journal, later supported by a 2389 paper and Anthropic research; Cialdini persuasion principles, later reproduced against frontier models by Cialdini, Mollick, and Shapiro — claims cited, not detailed).

Assessment

Freshest entry on this shelf (two weeks old at ingestion) and the only one addressing identity as an evolution locus — a surface the skills/rules/memory taxonomy doesn’t name.

• Durable: the receipt — a claimed lesson counts only when it lands in a versioned, inspectable surface through the designated writer; “I’ll remember that” is a red flag, a diff is evidence. That is the transferable idea for the evolve skill.
• Era-bound and unverified: single anecdote, no eval harness yet (said plainly in the post), product pre-release, author’s stake obvious.
• In this library: convergent evolution with the academic line — sole write access through a dedicated role is zhang2025‘s LLM-proposes/gatekeeper-merges split and zhang2026‘s proposer/promoter separation, applied to persona; the planned multi-day slow-roll adds a gate dimension (time) none of the papers use. It is also the direct answer to the “misevolution” and drift risks gao2025 catalogs — but where the survey’s checklist assumes human approval gates, this pattern deliberately replaces them with structural internal ones, a genuine philosophical fork worth watching.