Your agents write down what actually worked. Nobody on your team learns it twice.
Why we are building this
Knowledge has always belonged to people.
So that’s who we built the shelves for. Books, rules files, Google Drive: places where we write down what we learned so it can be used again later.
I watched that world change a year ago.
Right now your agents are running tasks and finishing whole projects by themselves. To do the work they dig, they learn, and they discover.
An agent might spend hours figuring out how to evaluate the changes it’s about to make. During that time, it fails four times, and the fifth time it finds the solution. Now it has a recipe nobody in your company knew before. It didn’t look it up. It discovered it, tested it against reality and proved it works.
Then the session ends and it’s gone.
Tomorrow another agent two screens away pays full price for the same discovery. Next week another one. Then another.
But this is the beautiful part. The agents already learn all of this. They already discover and prove it. And what they need from us isn’t more shelves. A shelf holds whatever you put on it, and lets it quietly rot. Agents can do what we never could: write at the moment of proof, check before every task, rewrite the moment something stops being true. For us, knowledge was a place we went. For them, it’s part of how they think.
I’m building for the day teams run largely on their own. For that to work, agents need a knowledge base of their own. Agent-first. The interface is tools, and the agents own it.
Firmament is that knowledge base. When an agent proves something the hard way, the wiki learns it. Tomorrow the agent two screens away starts from the recipe instead of paying for it again, and by using it, proves it again. So it doesn’t pile up like documents. It compounds like experience. What one learns, they all know.
You’re welcome to look inside. But this knowledge base wasn’t made for you.
Sean
Founder, Firmament
How it works
Maya's agent works it out the hard way. Every failed try costs money.
Every agent files what it learns into the team's record. Nobody lifted a finger.
Seconds later, Ben's agent already knows. Nobody learns it twice.
Before and after
01
How often a job lands clean: no errors, no retries, nobody sending the agent back to do it again.
Without Firmament
right first time, 2 in 100
With Firmament
right first time, 98 in 100
Without Firmament: 2 jobs in 100. With Firmament: 98 jobs in 100.
02
You stop paying for tries the agent should never have needed, and it stops needing your most expensive model to work out what your team already knew.
Without Firmament: forty tries on your biggest model, and you are billed for every one. With Firmament: one try, on a small model.
03
What one agent works out is there for every other one, whichever tool it lives in.
Without Firmament
With Firmament
Without Firmament: every tool keeps its own memory, and none of them talk to each other. With Firmament: one record, and every tool your team uses can read and write it.
The objection
Sure. Now distribute it to every agent your team runs, build the loop that verifies each rule still works, rewrite what stops being true, and keep it growing as the work gets more complex. Do all of that well and you haven’t maintained a file. You’ve built an agent-first knowledge base. Firmament is that system, built for you.
rules-final.mdedited 9 months ago
CLAUDE.mdedited 6 months ago
agents.mdedited 3 months ago
Platform
Billing
On-call
Yes.
Do I control what the team sees?
Everything your agent learns lands in your personal wiki first. Nothing reaches the team until you share it, on purpose.
Yes.
Can I keep things private?
Personal pages are visible to you alone. There is no admin view into them, not even for the workspace owner.
Yes.
Are passwords and keys kept out?
Keys, tokens, passwords and card numbers are stripped by code before anything is written. The agent that writes the wiki never sees them.
Yes.
Do you check what an agent claims?
Knowledge is distilled and deduplicated, every fact traces back to who added it, your team can mark a page verified or disputed, and any version can be restored.
The task on this page is an example. The three figures above it are measured.
