Headlong
ecosystem:headlong
Laude Institute's Apache-2.0 agent microharness (<10K lines of Bash) built around shellm, a recursive-language-model (RLM) core, for persistent agents that keep thinking between messages.
Specification
| Category | Ecosystem watch |
|---|---|
| Runtime | external |
| Interface | Named identity CLI |
| Source | industry watch |
| Licence | Apache-2.0 |
| Continuity | Owned by the agent |
| Support | Integration pending |
| Based on | Headlong (Bash, Apache-2.0) |
| Last verified | 2026-08-14 |
Setup and support
Not runnable from SuperQode, and a different case from jcode's: Headlong has no ACP server, MCP server, SDK, or task-runner API. `<agent> hello` is a one-shot observation sent into a standing mind that may choose not to reply, not a task call with a result contract. It has a Unix surface to attach to (start/stop the identity, tail its trajectory) but no run contract to speak of yet. See RLM Routes Compared for how its persistent inner-monologue loop differs from SuperQode's own RLM routes.
Runtime behaviour
Capabilities
Tools
thinkers (idle-loop dispatcher)shellm (Bash RLM core)traj (jsonl DAG, fork/merge)context (tiered-rollup projection)mem / skillsrecap (logarithmic memory pyramid)Policies and ownership
- +Headlong owns its agent loop, trajectory, and identity store
- +Model provider keys (Anthropic, OpenAI, Gemini, OpenRouter) stay in Headlong's own config
- +No commands are executed by SuperQode until a connector is built and tested
Install and authenticate
- 1
Install Headlong independently
Interviews you for an agent name and personality, then installs it as a command.
curl -fsSL https://headlong.ai/install.sh | bash - 2
Configure a model provider
Headlong's `llm` CLI takes an Anthropic, OpenAI, Gemini, or OpenRouter key.
