A conversational agent that drives the platform and “autonomously monitors and improves your agentic setup”, running alongside your existing coding agents (open beta). Invoked as tessl agent — interactive, one-shot, non-interactive with JSON output, resumable, model-selectable, with its own MCP server management (agent mcp add/get/list/remove) and a yolo mode that auto-approves permissions.
Its subject is the toolchain, not the codebase, which is what makes it unusual enough to page. The documented uses are things like mining PRs and session logs for recurring agent mistakes, turning those into a skill fix, and proving the fix with an eval; and spotting repetitive work in the development loop and turning it into an automated workflow. That is pattern-knowledge-compounding with the loop closed by a separate agent whose job is the agent setup — the framework-side instinct of pstack-reflect and ce-compound, lifted out of the session and made continuous.
It also completes a circle worth noticing. tessl-observability data — where skills actually activate across sessions, not merely where they are installed — is what this agent mines, and the fixes it proposes are published back through tessl-plugin and measured by tessl-eval. Observe, diagnose, fix, prove, redistribute: the compounding loop the process layer runs on a developer’s session, run instead on the organization’s capability corpus.
Filed as sub-agent because that is what it is in the wiki’s vocabulary — a delegated agent with a bounded remit — though it is a first-class product surface rather than something a command spawns.
See Also
- tessl-eval — how its proposed improvements get proven.
- tessl-inventory — the estate it reasons over.
- pstack-reflect · ce-compound — the in-session, framework-side counterparts.
- pattern-knowledge-compounding — the loop, run on the toolchain.