Scheduling, supervision, and memory for agents that run unattended
The problem
Agents that run on a schedule fail silently. The job stops, the log stops updating, and nobody notices for a week. Meanwhile the reasoning behind hundreds of decisions evaporates the moment each session ends.
What we built
Four adapters read sessions from Claude Code, Codex, Grok, and opencode, classify each one to a project, summarize it through a cheap headless model, and file the result into a vault with per-project indexes.
The digester spawns model sessions, and those sessions land in the very stores it reads, which turned one backfill run into a self-feeding loop that doubled its own workload. A sentinel filter now excludes any session containing the digest instruction.
An hourly monitor checks log freshness, process liveness, and an HTTP health endpoint across every always-on job, then notifies only when the failure set changes. An alert that fires every hour is an alert nobody reads.
A local proxy, a knowledge-graph server, and a response-compression mode run across all four agent CLIs, so long sessions stay inside a context window and cost less to run.
What this demonstrates
Operating agents in production: scheduling, supervision, cost control, and institutional memory.
Get in touch
Tell us what you are trying to build and what is in the way. We will tell you early if we are not the right fit.