JURNITI°
Login
All guides
Headlong logo
Headlong

Headlong on a MicroVM: The Always-Thinking Agent

Headlong is a Laude/MIT Bash microharness with persistent agency. Leave the mind running on a Firecracker guest. BYOK. About $1–2/hour while it thinks.

Most agent harnesses wait. You send a task. They finish. They freeze until the next request.

Headlong does not wait.

A message from you is not a new session. It lands in a thought stream that is already moving. The agent decides if and when to reply. Between messages it keeps thinking — debugging its own code, starting projects, pinging you when it has something to say.

That is persistent agency. Laude Institute and MIT shipped it as a Bash microharness (under 10K lines, "of bash, by bash, for bash") in August 2026. It is alpha research software. It is also the first harness in this catalogue whose default posture is unattended spend.

Run Headlong on a jurniti microVM

The laptop problem, multiplied

A local Headlong install is a curl. Leaving it running is the hard part.

  • Close the lid and the inner monologue dies with it.
  • Every thought is a billed model call. On a laptop that also holds your password manager, an agent with a shell is the wrong blast radius.
  • You want the identity, the memories, and the trajectory DAG to still be there tomorrow.

Headlong's own docs say: use a dedicated, spend-capped API key, and run it in a sandbox because the agent will execute shell commands. That is not a disclaimer to skip. That is the product.

What we actually run

jurniti gives Headlong its own computer — a Firecracker microVM, the same isolation primitive AWS uses for Lambda. Not a container. Not a shared box.

On provision you get:

  • A fresh Ubuntu microVM with headlong-init and persona already on PATH. The checkout is pre-baked.
  • A live in-browser terminal. Run headlong-init with YOUR key. The generic dashboard unit is masked; Headlong's own dash exists only after an identity exists, and you start it with persona dash.
  • Persistent ~/.headlong — identity, memories, trajectories, both key-bearing dotenvs.
  • Pinned CPU and RAM.

BYOK throughout. We never proxy model spend.

The spend you must look at before Enter

headlong-init ends by starting the mind. From that moment the loop is a model call on your key, including while you sleep.

Published range from people actually running it: about $1–2 per hour. Left up 24/7 that is about $720–1,440 per month, on top of the $25 microVM. A disclosure without those numbers is theatre.

Do this first:

  1. Create a dedicated provider key.
  2. Put a hard cap on it (OpenRouter: Settings → Credits → limit; Anthropic/OpenAI: per-key budget).
  3. Then run headlong-init.
  4. After the first day, read the bill. persona stop is the off switch.

If that range is a problem, this is the wrong harness. Use a turn-taking coding CLI.

Snapshot it — later

Public templates for Headlong are vetoed until a live capture canary has proven the scrub. Private restore keeps the identity; a public fork must not inherit your key, your chat, or your trajectories. We will not flip that veto on a spec we have not run.

Honest limits

  • Alpha upstream. No tags, no releases. We pin a commit. Expect the tree to move.
  • Continuous spend. This is not Claude Code sitting idle. The mind is the product, and the mind costs money.
  • No free tier. From $25/mo for the box. The de-risk is a 30-day money-back guarantee on your first purchase, not a trial.

Get the box

Pick a plan, choose Headlong, pay. The VM is up in a few minutes. Then:

# cap the key first, then:
headlong-init
persona status

Frequently asked questions

What is Headlong?
Headlong is an open-source agent microharness from Laude Institute and MIT. The core is under 10K lines of Bash. Its defining feature is persistent agency: the agent keeps thinking between messages in a self-guided inner-monologue loop.
Does it spend money when I am not at the keyboard?
Yes. Once `headlong-init` starts the mind, every thought is a model call on YOUR provider key. Measured range is about $1–2/hour (~$720–1,440/month if left running 24/7). Set a hard spend cap on the key before you bootstrap, and run `persona stop` to halt it.
How do I authenticate?
BYOK only. `headlong-init` asks for your OpenRouter, Anthropic, OpenAI, or Gemini key and interviews you to create the agent's identity. There is no subscription login.
Pricing?
From $25/mo for the microVM, BYOK for model spend, 30-day money-back on your first purchase. No free trial and no free tier. Model spend is extra, on your key.