Blog
Notes from the board
Writing about the parts of software delivery that happen after the code exists — review, isolation, deployment, and the quiet hour afterwards. Written by the people who build and operate Capybari.
Recent
Latest posts
Why every task gets its own git worktree
A branch on its own is not enough once four specialists are working at the same time. What a shared checkout does to parallel work, and why abandoning a task should mean deleting a directory rather than unpicking an afternoon.
ReadMoving through the three autonomy levels
Manual to Supervised is a decision about plans. Supervised to Autonomous is a decision about gates. They are different questions, and treating them as one is why teams stall halfway.
ReadWhat an agent should do when it doesn't know
Guessing is cheap to do and expensive to own. The rule lives in the role contracts: hit genuine ambiguity, stop the run, ask a question. The hard part is defining genuine.
ReadThe economics of an agent run
A run is what we meter because a run is what costs us money. Opening one up — plan, execute, review, report — explains why a small request sometimes costs more than a large one.
ReadWhy we publish what we cannot do
Our pricing page lists, by name, the things an enterprise needs that we do not have yet. Writing that list was uncomfortable. We would write it again.
ReadEngineering
How the work is isolated, ordered and reassigned
Posts about the machinery: worktrees, dependency graphs, parallel execution and what happens when a role is missing.
What breaks when four agents work in parallel
Isolation solves collisions in the working tree. It does not solve two tasks that both decide to rename the same function, so the interesting failures move up into the plan.
ReadDependency graphs make better plans than checklists
A numbered list records the order you imagined. A graph records what genuinely has to wait for what. The difference shows the moment one task blocks and three others could have carried on.
ReadWhy capability substitution beats failing a task
When a project has no Database Engineer, the work is reassigned to whoever can cover it — and the substitution is recorded. The record matters more than the reassignment does.
ReadOperations
Deploying, checking, and the hour after it goes red
Posts about deployment targets, health checks, rollback and the investigation that follows a bad night.
Designing a rollback you can trust
You write the rollback command, so it is only as good as the last time you ran it on purpose. Run it on a Tuesday afternoon, not during the deployment that needs it.
ReadThe shape of an investigation write-up
Timeline, cause, blast radius, what changed, what we still do not know. The last section is the one everyone skips, and it is the reason the next investigation is shorter.
ReadWhat a deployment target needs before you trust it
A deploy path, a deploy command, a health check, a diagnostics command, a rollback command. Only the health check can tell you the deploy went wrong, so write one that fails when it should.
ReadWorking with agents
Reading what they write, and writing what they read
Posts about the two documents you will spend the most time with: the request going in and the report coming back.
Reading a completion report
Summary, files modified, risks, assumptions, recommendations, blockers. Read the assumptions first — that section is where a task quietly tells you it filled in a gap.
ReadWriting a request that plans cleanly
Describe the outcome and the constraint, not the implementation. The worst plans come back from requests that had already decided which file to change.
ReadApproving a plan you did not write
You are not marking prose. Check the owners, check the order, then check what the plan has decided not to do. The third one is where the surprises live.
ReadBrowse
By category
Five things we write about, and nothing else.
Engineering
Worktrees, branches, dependency graphs, parallel execution and the failures that only appear at four tasks at once.
Operations
Deployment targets, health checks, rollback commands and the investigation that opens when a check goes red.
Autonomy
Manual, Supervised, Autonomous: what each level actually changes, and how to tell when you are ready for the next one.
Working with agents
Writing requests, approving plans, reading structured reports, and answering the questions a run stops to ask.
Company
How we decide what to build, what we refuse to claim, and why the awkward parts go on the website rather than in a sales call.
Changelog
Not an essay category. If you want the dated list of what changed in the product, it lives on the changelog.
The digest
One email a month
New posts, plus a short note on what shipped. No drip sequence, no sales follow-up, and one click to stop.
Reading about it is slower than watching it
Connect a repository, file one request, and read the plan that comes back. Manual autonomy — nothing runs until you say so.
No card required · Free plan forever · Export everything, any time