Headlong microharness in SuperQode Ecosystem Watch

Laude Institute recently released Headlong: Apache-2.0, under 10K lines of Bash, with shellm as a recursive language model core. A Headlong agent keeps thinking between messages. A Slack or Telegram line is one observation on a single thought stream; the agent decides whether, and when, to answer. SuperQode indexes it under Ecosystem watch as ecosystem:headlong in :hub and on the Harness Hub. Headlong owns the mind. SuperQode lists the contract.

How Headlong works

Most coding harnesses wait for a task, finish a run, then sit idle. Headlong's answer is persistent agency: an inner monologue with exponential backoff while the channel is quiet (five seconds, ten, twenty, doubling toward a cap) and an immediate reset on a new message. Background thinking costs about one to two dollars an hour at Laude's published settings. The model writes bash. thinkers chooses the next thought. shellm runs the block until FINAL is set, and nested calls fork child branches. Generated code uses Docker when Docker is available. traj is an append-only jsonl DAG with fork and merge; context and recap project that life into the next prompt as a logarithmic staircase. The name becomes the command: ada hello, ada, ada dash.

Laude's internal agent, Audel, has had more than fifty of its own commits merged into main, including an unprompted forty-eight-minute repair of a recall process it had built and left unwired. A team shares one identity, so the timeline is shared too. Docker is the default sandbox for generated code; Audel itself runs on a dedicated VM, which is Laude's operational choice.

Headlong and SuperQode's RLM routes

Recursive language model is an inference strategy from the October 2025 RLM paper: context lives as a REPL variable, and the root model writes code that greps, chunks, and calls sub-models. SuperQode already runs three routes on that line. Native RLM is first-party (one python tool, resident root, Host / Docker / Monty). RLM Code is an optional package (paper semantics, LID, bounded scored runs). Prime Agent is a connected product over RPC and ACP (IPython kernel, live child sessions). Headlong takes the same recursive-call idea, swaps Python for bash, and wraps it in a standing mind. The paper's author called that continuous thought the property that sets this harness apart. Prime's daemon persists between prompts. Headlong's mind keeps generating thoughts between them.

Why SuperQode cannot wrap Headlong

Runnable Hub rows speak a surface SuperQode already implements: ACP for fx, Prime Agent, OpenHands, Hermes, and Pi; JSON-RPC for DeepSeek Harness; an in-process backend for RLM Code; a first-party kernel for Native RLM. Those adapters all expect a completed run, a session, or a backend spec. Headlong publishes Unix identity commands (start, stop, hello) and jsonl on disk. Mapping that onto run.completed would treat an observation as a finished task. jcode at least ships a headless jcode run and an SDK. Headlong ships a mind that may stay silent. That mismatch is why the row is Ecosystem watch, and why wrapping Headlong like Codex or fx is the wrong shape.

We tried three attaches and each one misrepresented either containment or the job. Wrapping <agent> hello under SuperQode execution_policy sandboxes the one-shot process while the thinkers loop, chat bridges, and background mind keep running under Headlong's own trust model. Sandboxing the identity for its whole lifetime either kills the standing mind (the entire product) or invents a daemon sandbox SuperQode leaves unused for Hub connectors. Scoring the standing loop against RLM Code's twelve-step LID profile on harness bench compares an always-on team mind with a deny-by-default analysis call. Public discussion pressed the same gap: team-wide isolation on one trajectory, curl | bash for an alpha agent that runs shell around the clock, and a loop that spends tokens unless a review or training signal closes. The authors shipped before a persistent-agency benchmark and pointed at Terminal Bench 3 as a generic harness metric.

Headlong on Ecosystem watch

ecosystem:headlong is in the Hub with readiness integration pending and Apache-2.0 from GitHub metadata. Start, stop, and traj are the Unix surface. hello is an observation into a standing mind. Docs match on the Hub page and RLM Routes Compared. Next work, if we build it, is identity attach and trajectory import, leaving execution with Headlong. Porting observation-versus-session, thinker backoff, and tiered life-context onto Native RLM stays a research direction, unshipped.

Install Headlong yourself if you want the mind. Use a dedicated, spend-capped key. See the Hub docs, Headlong, the Laude launch post, and PyPI.