Laude Slingshots Four: 14 Open Research Projects, and the Headlong Agent That Never Sleeps

The Laude Institute has announced Slingshots // FOUR, its fourth batch of grant recipients: fourteen projects from eight universities and research institutes, spanning open models, self-improving agents, GPU kernels and research tools. The list, on Laude's announcement page, reads like a map of where open computer science research is heading. One entry is Laude's own: Headlong, a Bash agent harness that keeps thinking when nobody is talking to it, which Laude described in detail in a separate post.

Slingshots gives fast, low-friction grants and hands-on support to computer scientists turning research into startups or open source infrastructure. This guide lists all 14 projects as Laude describes them, then goes deep on Headlong, because it comes with real operating numbers and failure reports.

The 14 projects

Laude groups them into themes: openness (Marin publishes weights, data, logs and decisions from its hero run; Atlantic-CC shares the data curation strategy behind it), self-improvement (Pedagogical RL on training, SkyDiscover on search over algorithms and systems, ACE on context, Exoharness and dotagents on the harness), and "no piece of the stack underexplored" (kernels, serving, query engines, environments and retrieval). Laude says it accepts projects year-round and announces batches every few months, and that its support varies from compute to product strategy to a place to build.

These are Laude's descriptions at announcement; most projects will publish their own results later, and none of the performance claims above has been independently checked.

Headlong: an agent that never sleeps

Most agent harnesses are reactive: you give a task, the agent works until done, then sits frozen. Some add cron jobs or heartbeats that run a fixed checklist. Headlong works differently. Per Laude, "the agent is never asleep and there is no checklist unless the agent creates one." It generates thoughts about whatever it finds interesting in a self-guided loop, and a message from a human "doesn't start a session," it is "one more observation" in the thought stream.

How it is built

• Tiny core. Under 10K lines of Bash (9.9K lines in two directories). Tools, agent framework, memory and skills are executables and files, so the agent can inspect and modify any part of itself.
• The loop. A thinker repeatedly calls shellm, a Bash implementation of a recursive language model, which calls an LLM for reasoning text or a bash script to run, until it sets a final variable. A context tool assembles each prompt and a traj tool writes thoughts to the trajectory.
• Memory. A compaction scheme keeps the whole trajectory in context at exponentially decaying resolution (recent entries verbatim, older ones summarized), and the trajectory is a DAG of jsonl files with fork and merge.
• Skills are markdown files included in context; mem and traj ship by default.

What Laude reports from running it

Laude ran a shared agent named Audel on Slack, Telegram and a mobile app for weeks. Highlights, all from Laude's own write-up:

• Multi-player. One thought stream serves everyone, so there are no per-user sessions. Audel once reviewed two teammates' branches unprompted and caught a hardcoded model name, and on day one audited a teammate's eight stale branches, then corrected its own count ten minutes later. The flip side: Audel "is bad at keeping secrets," so assume anything you tell it is shared with the team.
• Self-repair. On August 5, Audel built a recall process, later checked on its own whether it was wired in, found it was not (it read an environment variable nothing set), diagnosed it, and fixed it in 48 minutes. Laude pulled the fix into main, and over 50 of the agent's commits from its own fork.
• What broke. A 30-second silence watchdog killed Audel's recursive sub-runs, and it fought the watchdog for about 40 minutes before mostly abandoning them (64 merges in two days, 12 in the next twelve). Three times it stopped its own service by accident, so Laude added a guard, which Audel itself later found was too broad.
• Cost. Thinking in the background costs $1 to $2 an hour at Laude's settings with GLM or Grok. The thinking rate backs off from 5 seconds to a configurable cap when idle and resets when a message arrives.
• Evaluation. Laude says improvement is measured qualitatively today and asks for ideas on measuring the long-term value of persistent agency.

Safety notes

Laude calls Headlong "alpha research software," says agents "can and will run shell commands," and recommends a sandbox and a dedicated spend-capped API key. With Docker installed, Headlong runs each bash block inside a container, exposing only what you mount and the credentials you hand over; Laude runs Audel unsandboxed on a dedicated VM, which is not the default. The install is a one-line script piped to bash, which you should read first.

Who might want to try Headlong

Headlong is aimed at researchers and tinkerers, not at teams that need a dependable assistant today. It makes sense if you want to study persistent agency (what happens when an agent has an inner monologue and picks its own projects), if you want a harness small enough to read in an afternoon and modify, or if you run a small team and want a shared agent that notices things across everyone's work. It makes less sense if you need audit trails per user, strict secrecy between colleagues, or predictable spend. The always-on loop means a budget cap is not optional, and the single shared thought stream means access control has to be handled outside the agent. Laude itself says it has not studied what happens when two people give conflicting instructions.

The design is also a useful reference even if you never install it. Treating tools, memory and skills as plain executables and files means an agent can read its own implementation; treating the trajectory as a branching log with a context window that is a projection of it is a clean way to think about memory; and decaying-resolution compaction is a concrete answer to the short-term memory problem that long-running agents hit. Whether those ideas hold up outside Laude's own team is the open question, and the institute is explicitly asking for outside experiments.

How it compares with other persistent agents

Laude's background section places Headlong next to MemGPT, OpenClaw, Hermes Agent, Prime Agent and Exo. Our own coverage of the neighbors: OpenClaw 2.0, Hermes Agent versus OpenClaw, xAI's Grok Bot persistent agents in early beta, and the LoopX agent control plane. The distinguishing idea is not scheduling but removing the idle state altogether.

Why this batch matters

Three trends show up. Openness is moving from weights to the whole recipe, with data, logs and decisions public. Self-improvement is being attacked at every layer, training, context, search and harness, and several of the projects share people and ideas (Alex Krentsel appears on both Exoharness and SkyDiscover; Noah Ziems and Omar Khattab on Pedagogical RL and Headlong). And the infrastructure layer, kernels and serving, is getting the same open treatment as models. For readers following our coverage of training methods, Perplexity's JOLT self-distillation is a related line of work on teaching models from their own attempts.

What to watch

• Papers and repositories for Pedagogical RL, ACE, Exoharness and SkyDiscover with measured results.
• Whether Marin and Atlantic-CC publish the full run artifacts as promised.
• Independent reports from people running Headlong, especially on cost, secrecy and safety.
• Which of the 14 become startups versus remain open infrastructure.

Related reading

• OpenClaw 2.0
• Hermes Agent vs OpenClaw
• Grok Bot persistent agents
• LoopX agent control plane
• What are agent skills?
• Official: Slingshots // FOUR and Headlong post

Details are accurate as of October 22, 2026 and come from the Laude Institute. Project descriptions and results are Laude's own.