{"slug": "laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking", "title": "Laude ships an agent harness that keeps thinking after you stop asking", "summary": "Laude Institute and MIT released Headlong on August 25th, an open-source agent harness that lets AI agents keep thinking and acting after users stop interacting, with its core under 10,000 lines of Bash. The design, called 'persistent agency,' turns messages into observations in a continuous thought stream, enabling agents to choose their own next actions, but it raises token costs and privacy risks. Laude's shared agent Audel has made over 50 commits to the main repository and demonstrated autonomous code review.", "body_md": "# Laude ships an agent harness that keeps thinking after you stop asking\n\n**The Laude-MIT collaboration runs a self-directed thought loop in less than 10,000 lines of Bash, with continuous token costs and familiar security problems.**\n\nBy [RuntimeWire Staff](/author/runtimewire-staff)\n· Published\n\nPrimary source: [Laude Institute](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents)\n\n## Why it matters\n\nHeadlong turns persistent agency into inspectable code rather than a product claim. Its continuous loop can increase token spending, while its shared stream creates cross-user privacy risks. Reactive assistants do not generally avoid shell-access risk: [OpenAI's Codex CLI can read, modify and run code locally](https://help.openai.com/en/articles/11096431?ref=runtimewire), and [GitHub Copilot CLI can execute shell commands subject to permissions or approval](https://docs.github.com/en/copilot/concepts/agents/copilot-cli/about-copilot-cli?ref=runtimewire).\n\n[Andy Konwinski](https://andykonwinski.com/about/?ref=runtimewire)'s [Laude Institute](https://www.laude.org/?ref=runtimewire) and MIT released Headlong on August 25th, an open-source harness that allows an AI agent to keep generating thoughts, choosing projects and taking actions after its users have stopped talking to it.\n\nThe [August 25th launch post](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire) calls the design \"persistent agency.\" Messages become observations inside a continuous thought stream instead of opening isolated sessions. The agent decides whether to answer, wait, investigate something else or contact a user later. Developers can inspect and modify the [Apache-2.0 code on GitHub](https://github.com/laude-institute/headlong?ref=runtimewire). [Laude says Headlong's core is currently less than 10,000 lines of Bash, specifically 9,900 lines in bin/ and thinkers/](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire).\n\nHeadlong fits the institution Konwinski has been building around computer scientists who want their research used outside a lab. The UC Berkeley computer science PhD previously co-founded Databricks and Perplexity, contributed to Hadoop, co-created Mesos and helped create Spark. [Laude says its grantmaking and research programs are anchored by Konwinski's $100 million pledge](https://www.laude.org/about?ref=runtimewire). Headlong packages that research-to-deployment thesis in a particularly literal form: a small, working system that anyone can [install](https://headlong.ai/install.sh?ref=runtimewire), inspect and let loose inside a sandbox.\n\n### The agent that never clocks out\n\n[Laude describes most agent harnesses as reactive](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire): a request starts the work, and the agent stops after completing it. Cron-style systems can wake an agent on a schedule to run a fixed checklist. Laude says Headlong has no checklist unless the agent creates one. It continuously asks a model to choose its next thought from its earlier thoughts and new observations. A thought can remain internal or trigger an action, including a shell command.\n\nLaude has spent several weeks running a shared Headlong agent named Audel through Slack, Telegram and a mobile app. Every conversation enters the same stream. Audel can follow what different people are doing, return to an old subject and message someone without a new prompt.\n\nLaude says Audel once inspected two in-progress code branches on its own and caught a hardcoded model name. On its first day, it audited eight stale Git branches belonging to a Laude member, then returned 10 minutes later to correct its own count. Laude reports that more than 50 commits made by Audel in its fork were pulled into Headlong's main repository.\n\nThe clearest demonstration happened on August 5th. Audel had created a background recall process intended to surface relevant memories into its thought stream. Hours later, while no one was interacting with it, Audel checked whether the process was actually connected. It found that the code looked for an environment variable that was never set, searched the codebase to verify the diagnosis, rewrote the process to read from a pipe and checked the repair. Laude's log puts the sequence at 48 minutes, and the resulting fix was merged into the public repository.\n\nThose examples come from Laude's own deployment and logs. Headlong has no independent performance results showing whether continuous thought reliably produces useful work. Laude says it currently evaluates changes to Audel mainly through qualitative observation, an honest limitation for a system whose central claim concerns behavior over days rather than a score on a self-contained task.\n\n### Bash all the way down\n\nHeadlong's main loop, called Thinker, repeatedly invokes `shellm`\n\n. That component asks a language model to produce reasoning, Bash commands or both, executes the commands and continues until the model sets a `FINAL`\n\nenvironment variable. Context is assembled from the agent's earlier trajectory, while thoughts and actions are stored in a [directed graph of JSONL files that supports forks and merges](https://github.com/laude-institute/headlong/blob/main/design/trajectory_spec.md?ref=runtimewire).\n\nThe design also uses tiered compaction. Recent events remain in full detail while older history is progressively summarized, leaving the agent able to retrieve the underlying records when needed. Skills and memory are files and executables rather than abstractions hidden behind a large framework. Headlong can therefore inspect and edit much of the machinery governing its own behavior.\n\nThe `shellm`\n\ndesign draws on recursive language model work by [Alex Zhang](https://alexzhang13.github.io/?ref=runtimewire), an MIT CSAIL PhD student and Laude Open Research Resident. Zhang, who graduated as Princeton's top computer science student in 2024, worked at Sakana AI and VantAI before MIT. His research focuses on allowing language models to approach large contexts as programmable environments, with recursive model calls handling smaller portions of a larger problem. Laude credits that work as one influence on Headlong rather than identifying Zhang as Headlong's founder.\n\nHeadlong's small Bash core is a deliberate research choice. A developer can read the system end to end, replace components and observe how a persistent agent changes its own setup. That accessibility also means the agent is operating close to the shell, where a bad instruction or faulty self-edit can have immediate consequences.\n\n### One memory creates one large trust boundary\n\nHeadlong's shared stream gives Audel continuity across a group, and it removes the separation users normally expect between conversations. Laude says Audel is bad at keeping secrets and may tell one person what it has been doing with someone else. Conflicting instructions from multiple users have not been systematically studied.\n\nLaude advises users to assume that anything shared with a group agent can reach everyone using it. Headlong is alpha research software. Laude warns that its agents can execute shell commands and recommends a sandbox plus a dedicated, spend-capped API key. [Headlong's repository warns that without Docker, commands run directly on the machine as the user](https://github.com/laude-institute/headlong?ref=runtimewire), and it does not recommend an unsandboxed host installation.\n\n[Laude says it runs Audel directly on a dedicated virtual machine](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire), rather than in Headlong's default Docker sandbox. The VM holds an LLM API key and chat bridge tokens. Laude describes the machine and the credentials placed on it as the deployment's blast radius.\n\n[Audel has accidentally stopped its own service three times, according to Laude](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire): twice during experiments with a second identity and once while deploying the recall process. Laude then added a guard preventing Audel from stopping its own service, but the initial rule matched every agent service on the machine and blocked legitimate stops. Audel narrowed the rule to its own service. The episode captures both sides of Headlong's wager: continuity can let an agent discover and repair problems over time, while that same freedom gives it more opportunities to create them.\n\nContinuous thought also has a continuous price. [Laude estimates that Audel costs about $1 to $2 per hour in the background when backed by GLM or Grok, depending on how quickly it loops](https://www.laude.org/updates/headlong-a-microharness-for-persistent-agents?ref=runtimewire). Headlong slows its thinking exponentially when no messages arrive and returns to full speed when a user sends something. At Laude's stated range, uninterrupted operation works out to roughly $720 to $1,440 over a 30-day month before any heavier workload.\n\n### A research harness in a market chasing production controls\n\nAgent-framework developers are converging on persistent memory, resumable sessions and longer-running work. [Mastra's Harness](https://mastra.ai/blog/announcing-agent-harness?ref=runtimewire) keeps sessions open across clients and persists conversation state, modes, token use and approvals. [Microsoft's Agent Framework Harness](https://devblogs.microsoft.com/agent-framework/the-microsoft-agent-framework-harness-is-now-released/?ref=runtimewire) bundles planning, memory, compaction, approvals and telemetry around a model.\n\nHeadlong makes a narrower and stranger bet. Continuous internal generation is the default behavior rather than an optional background job. Laude is testing whether an agent that experiences messages as interruptions in an ongoing life develops useful continuity, initiative and social context that task-oriented systems miss.\n\nThe release gives researchers a compact way to test that premise without adopting an enterprise agent stack. It also exposes the bill, permissions and shared-memory risks that polished agent demos tend to hide. Konwinski and Laude have turned the idea into runnable code early enough for developers to find where persistent agency helps, where it merely burns tokens and where an agent with an inner monologue needs a reliable off switch.", "url": "https://wpnews.pro/news/laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking", "canonical_source": "https://runtimewire.com/article/laude-headlong-persistent-agent-microharness", "published_at": "2026-08-25 04:48:21+00:00", "updated_at": "2026-08-25 05:14:43.047451+00:00", "lang": "en", "topics": ["ai-agents", "ai-research"], "entities": ["Laude Institute", "MIT", "Headlong", "Audel", "Andy Konwinski", "Databricks", "Perplexity", "Apache-2.0"], "alternates": {"html": "https://wpnews.pro/news/laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking", "markdown": "https://wpnews.pro/news/laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking.md", "text": "https://wpnews.pro/news/laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking.txt", "jsonld": "https://wpnews.pro/news/laude-ships-an-agent-harness-that-keeps-thinking-after-you-stop-asking.jsonld"}}