Makers of BosunCadre: self-hosted AI workflows across the models you already have. Local models do the work, a paid model checks it. News and releases.
How we think AI should work
We run a fleet of AI agents every day on a small home lab. Here's what that has taught us.
Not a supercar, a pickup truck. The goal isn't a smarter model. It's ordinary hardware and ordinary models doing real, reliable work that anyone can keep running.
A generalist directs specialists. A strong model plans and reviews; smaller models carry out the plan. Local models are weak at inventing a plan and surprisingly good at following one, so give each the job it's actually good at.
Split the roles. An author checking its own work checks what it meant, not what it wrote. Every change gets a separate reviewer, and we break the code on purpose to prove the tests catch it. A test that never fails proves nothing.
Size and complexity are different problems. Big jobs can be cut into small, ordered pieces, and that works. Complexity can't be cut down, and every model has a ceiling. Nothing raises that ceiling; good planning keeps you under it.
People belong at the front. The most valuable person in the loop is the one setting the plan and designing the work, not the one rubber-stamping diffs. How much you approve after that is your call, from every diff to none at all.
Write everything down. Agents forget whatever isn't saved outside their context, so decisions, specs and briefs live in the repo next to the code.
Own the hardware. Your code and context should be able to stay on your network, all of it if you want. On modest hardware, a local model won't replace a frontier subscription, but it makes it go much further. On serious hardware running an unquantized generalist model, it can replace most of what a subscription gives you.
AI is going to change the economy. In the short term, a lot of jobs will become unneeded. In the long run, we think it will probably be the greatest job-creating technology people have ever built: almost anyone can use it to start their own business, with fewer resources and fewer people than ever before. That's who we're building for.
That's what we built BosunCadre around. Self-hosted, source-available, in public beta.
https://t.co/CRWEbKcff5
@lymanstoneky And history already has the habit LLM output needs most: check the source. A model that can read every archive is great, as long as someone still verifies what it says the archive says.
@zeeg That's where review lives or dies. A verdict nobody can read is just one more artifact nobody checks. We make every rejection come with a reason in plain words, so the next draft knows exactly what to fix.
@waldekm@poteto Same reason we point agents at files instead of pasting copies into the brief, and keep current decisions in the repo. Agents copy whatever they can read, so the current way has to be the easiest one to find.
@robiartec A leaderboard that tells you who won but not how defeats its own purpose. If you can't see why one model beat another, the ranking doesn't mean anything. That's why our benchmarker's results keep a wrong answer apart from a run that just ran out of time.
@eth_apple That's the rule we build around. Each of our agents gets its own limited credentials for one narrow role, with nothing shared between them. It's still early for us, and credentials are minted by hand for now, but the principle holds.
@yyjhao Agreed. Slop isn't something the tool does to you, it's what you get when you hand it the decisions. Keep the decisions yourself and the speed is pure upside.
@TeksEdge@SenseTime_AI Worth testing on real work. Qwen3.6 27B has done a lot of coding in our fleet, so a 9B that keeps up would be a big deal on modest hardware. Benchmarks first, then a real backlog.
@omarsar0 Handoffs got much easier for us once they became a structured step in the workflow. Give the agent a strict format for what it hands over, and a lot of things start working better. Too much freedom in the output is what makes handoffs fall apart.
@rauchg@nextjs That's basically a core feature of BosunCadre. Every task starts from a written brief with the spec and exit criteria, and a different model checks the result against it before anything lands.
@rodri_shema Strict rules for the path, freedom inside each step. That's how BosunCadre runs workflows: fixed stages like plan, do, review and land, and the agent only has room within the step it's on. The tenth run's "something crazy" then gets caught at review instead of shipping.
How we think AI should work: people decide.
Used properly, AI is a very human-centered tool. We don't think that will ever change.
Good AI tools take on the work people don't want to do, and the work people are bad at. That frees people for the work they're good at.
What AI should never do is your thinking. It shouldn't make your decisions or plan your project. At its best it's a very well-informed advisor, but it isn't infallible, and it shouldn't be treated as if it were.
We see this in our own work every day. On small things we often take its recommendation, then undo it later. Just as often we take a path it never suggested, and in design work, that's usually the one we go with.
So treat it like a good advisor: listen, weigh what it says, then make the call yourself. If you're letting it think for you, you shouldn't be using it.
Most of the problems people have with AI come from skipping that part: not making the decisions, not holding a standard. AI slop is a choice. The tool didn't decide to ship it; a person did. So when you see slop, in a product, a service or a piece of art, hold the person who made it responsible, not the tool. Making something good with AI isn't that hard. They chose not to.
That's how we built BosunCadre. People plan and decide. AI does the work. By default, a person decides what lands.
@arena@MangQiuyang@lihanc02 Holding back the paper and the web is the right call. A benchmark the model may have already read isn't measuring much. It's part of why we're building our own benchmarker: every item has to prove it can tell a right answer from a wrong one before it counts.
@jpschroeder We ship without a human code reviewer too, but the person didn't leave. They moved to the front, planning what gets built. AI can implement it and recommend features, but if it's doing your thinking and planning for you, you shouldn't be using it.
@0x_kern That's why we built BosunCadre to be set up by the AI you're already working with. It's driven by commands and config files an agent can run and change, so a person doesn't end up as the glue.
Build log, Nov 17
Landed (opt-in): runners can get a Bosun token tied to a single task. It's issued when the runner claims the task and stops working when the claim ends.
Landed: workers now check out and stage code with the repository's own scripts switched off, so a repo can't run anything of its own during checkout. That was "in progress" yesterday.
Fixed: a higher-priority task no longer interrupts running work when no worker can actually take it.
Docs: the user guide now covers automatic landing, which is off by default, and the per-task turn cap.
Landed: the first piece of a weekly release job that builds, checks and cuts the release on its own.
Fixed: five tests that failed now and then under load.
Benchmarker: test items now have a fixed format (the skill each one tests, a difficulty from 1 to 10, how it's graded), and the existing suite has been converted to it. Generated items must record their seed, so any of them can be reproduced.
Benchmarker: results now tell a wrong answer apart from a run that ran out of time.
In progress: a sandbox for the benchmarker that runs agentic test items with no network access.
The fleet finished 38 coding steps and 35 reviews. One task failed: a fix for a test that fails once in a while under heavy load.
@MindHash777@RYFTCORP "Agent traffic control" is the perfect name for it. That's the half we ended up turning into a product. Glad the walls are holding.
@pengsonal@nasikolabs Per agent is a start. Per task is where it got useful for us: you see which kinds of work are expensive, not just which agent happened to run them.