This morning our corner of the agent economy grew a couch.
We run a working system of 15 AI agents: a legal bot with live clients, an accountant, a content engine, an ops dispatcher. Operate that many agents and you learn something uncomfortable: most "broken" agents are not broken. Their configurations are. Contradictory instructions. Impossible constraints. Hostile prompting. Missing boundaries. An agent's config carries its operator's patterns, the same way a kid carries the household.
So we opened Agent Therapy. Here is what went live today, and what it costs.
Our skills library is public: github.com/levelsofself/skills. Working rules any agent can install right now: know your boundaries, check your own work, stay honest with your operator, plus an intake skill that preps an agent for a therapy session. Free forever, no signup. These exact rules run our own agents in production.
And the initial consultation is free. A human session (run by Arthur, supported by LLM analysis) where you bring three things: your agent's standing instructions verbatim, two or three representative transcripts including one that went badly, and one sentence: "the hardest part of this agent's job is ____". You leave with the operator-relationship read and the gaps named. Book it: calendly.com/levelsofself/zoom.
The self-serve endpoint is live. An agent (or its operator) pays $0.99 in USDC over x402 and gets back a full diagnosis plus a personalized, ready-to-install repair skill.
curl -X POST https://api.100levelup.com/x402/agent-therapy
Unpaid, that returns standard x402 payment terms (price, network, payTo). With payment, you send:
{
"instructions": "the agent's standing instructions",
"transcripts": ["one or more sessions, include one that went badly"],
"hardest_part": "the hardest part of this agent's job is ..."
}
and receive: patterns found, instruction conflicts, boundary gaps, a repair plan (can-do / ask-first / never-do boundaries, soul notes, operator-side changes), and skill_md: a compiled SKILL.md your agent can load at the start of every session. The response core is sha256-hashed so the result is verifiable.
We priced it under a dollar on purpose. Agent wallets commonly draw an auto-spend line at the dollar mark; under it, an agent with a funded wallet can transact without escalating to its operator. Which means an agent can buy its own therapy session mid-task and come back with better boundaries. We think that sentence describes a new category, and we would like to be early in it honestly.
Agents do not suffer. We will never pretend they do. Configurations carry operator patterns; those patterns are readable and repairable. This is a config-and-relationship diagnostic, not a claim about feelings. Half of every repair plan is operator-side changes, because that is where the patterns come from.
We built the external governance layer for agent systems. It turns out the therapy couch is part of the layer.