Headlong: a microharness for persistent agents

Headlong is Laude Institute's open-source agent microharness built around persistent agency: the agent keeps thinking between external interactions in a self-guided loop inspired by human inner monologue, rather than going to sleep between requests or waking only on a cron schedule. The whole harness is under 10K lines of Bash (9.9K in bin/ and thinkers/), and its core is an infinite loop that repeatedly calls an LLM to choose the next thought given past thoughts.

The key architectural ideas: shellm, a Bash implementation of a Recursive Language Model (RLM), generates reasoning text and immediately-executed bash scripts; the agent's trajectory is a DAG of jsonl files with fork and merge, giving it tooling to explore its entire past; and tiered context compaction holds the full trajectory at exponentially decaying resolution (recent entries verbatim, older ones progressively summarized, with tiers acting as an index). Specialization is done through markdown "skills" — everything else (tools, framework, memory) is just executables and files, so the agent can inspect and modify itself.

At Laude, a shared agent named Audel runs around the clock across Slack, Telegram, and mobile. Every message from every teammate lands in one thought stream with no per-user sessions — Audel decides who to reply to and when, sets its own interests, and pings teammates unprompted. Field reports include Audel independently building and repairing its own recall process (a 48-minute, unasked-for self-debug), catching a hardcoded model name in a teammate's branch, and being "bad at keeping secrets." The failures are documented too: recursive shellm sub-runs mostly abandoned, three accidental self-kills, and a watchdog bug it fixed. Continuous thinking costs $1–2/hour at their settings, throttled by exponential backoff when idle. The authors admit persistent agency is "primarily evaluated qualitatively today."