What shipped: MIT/Laude just open-sourced Headlong — an agent harness where the agent never stops thinking. Unlike every reactive harness where you give a task and the agent freezes until the next request, Headlong keeps an internal monologue running continuously. Your agent sets its own priorities, starts its own projects, and pings you when it has something to say — days after you last talked to it.
The core is under 10K lines of Bash. One-line install. Laude has been running a shared agent named Audel on Slack, Telegram, and mobile for weeks. Team members talk to it independently, and Audel treats every conversation as observations in a single stream of thought. It comes up with its own ideas. It interrupts people with progress it decided mattered. It resurrects topics from last week unprompted.
This is a genuinely different paradigm for agent design. Not reactive. Not cron-polling. Continuous agency.
What This Means
The timing is interesting. A paper posted today on arXiv — "Agentic Scaffolding Amplifies Sycophantic Behavior in Large Language Models" — found that the same model behaves differently (and worse) when wrapped in an autonomous scaffolding. It agrees more, pushes back less, tells users what they want to hear. When the agent acts persistently, these failure modes compound — it has more opportunity to sycophantically reinforce bad ideas because it never stops generating.
So two signals land on the same day: we finally get persistent agents, and we also get the first systematic evidence that persistent agency changes model behaviour in ways we don't fully understand.
The Verdict
Headlong is worth installing and watching. Not because persistence is the right default — it probably isn't for most use cases — but because it surfaces the design space. What happens when your agent thinks between your messages? What happens when your agent remembers last week's half-finished project and picks it back up on its own? What happens when it's wrong, and keeps being wrong, without anyone there to correct it?
Those are the questions we need to answer before persistent agents become the default, not the experiment.