I spotted Headlong doing the rounds on Hacker News this week — a "microharness for persistent agents." Small project, quiet announcement, but it's poking at something I've been thinking about for years: the difference between an AI that answers a question and an AI that has a job.
Most of what people build with LLMs is still the chatbot pattern. You ask, it answers, the process ends, memory gone. Fine for a lot of use cases. Useless for the ones I actually care about — content pipelines, monitoring systems, anything that needs to run at 3am without me watching it.
One-shot thinking doesn't survive contact with reality
The moment you want an agent to do something over hours, days, or indefinitely, a whole set of problems shows up that never mattered in demo-land:
- What happens when it crashes halfway through a task?
- How does it remember what it already did yesterday?
- What stops it wandering off-task after step 40 of an unsupervised loop?
- Who's watching it, and how do you know it's still doing the right thing?
None of this is exotic. It's the same set of problems every backend engineer has solved for cron jobs, message queues and long-running workers since long before "agent" was a buzzword. What's new is that the thing running inside the harness is non-deterministic, occasionally confident and wrong, and prone to drifting from its brief if you don't keep it on a lead. A microharness, as I understand the idea, is basically that lead — a small, opinionated scaffold that handles state, retries, task focus and lifecycle, so the agent itself can stay simple and swappable.
I like this pattern a lot more than the "give the model more tools and hope" school of agent design. Complexity should live in boring, well-tested infrastructure, not in the unpredictable bit.
This is exactly the problem I've been solving for years
I didn't call it a microharness when I built the pipelines behind RSSMasher, but that's what it is. Content ingestion, filtering, rewriting, publishing — that has to run continuously, recover from failures without human hand-holding, and not quietly go off the rails after the tenth iteration. The "harness" there isn't glamorous: logging, checkpoints, sane retry logic, and hard boundaries on what each stage is allowed to touch. The AI does the creative transmutation — raw feed into readable gold — but it does it inside guardrails that don't care how clever the model feels that day.
The lesson that translates directly to agentic AI: the model is not the product. The system around the model is the product. Anyone who's shipped automation tools for real customers over real months already knows this instinctively, because customers don't forgive silent failures. They notice when the pipeline stops, or worse, when it keeps running and quietly produces rubbish.
What this means if you're building with agents now
A few practical takeaways, from someone who's been burned by "it worked in the demo" more times than I'd like to admit:
Separate the loop from the logic. Your persistence, state management and task-tracking should be dumb, boring, and independently testable — not entangled with prompt engineering.
Design for graceful failure, not just success. An agent that can't tell you it's stuck is worse than one that stops.
Keep the agent's job narrow. The more scope you hand a long-running agent, the more surface area there is for drift. Microharnesses work because they constrain, not because they empower.
Watch it like you'd watch a junior developer, not a script. Long-running autonomy needs monitoring baked in from day one, not bolted on after the first embarrassing incident.
Where I think this goes
I expect microharness-style patterns to become as unremarkable as ORM libraries within a couple of years — infrastructure nobody talks about because everybody just uses one. That's a good outcome. The interesting work will move up a layer, into what these persistent agents actually get trusted to do unsupervised: research, content operations, monitoring, the unglamorous long-haul jobs that never fit neatly into a single prompt-response exchange.
If you're building anything that needs an AI to keep working after you've closed the laptop, this is the bit worth getting right. The model will keep getting smarter on its own. The harness is still your job.
— Wayne