We built Cogitorium because we were sick of the same fake choice every agent tool gives you: either trust one black-box model with everything, or duct-tape together scripts and pray you never have to explain the bill. So we didn't pick a side. We built a graph. Every agent gets its own model — throw a frontier model at the hard reasoning, let free local models grind through the boring stuff, all in one topology. Delegation only happens on edges you draw — not a rule in a doc somewhere, an actual permission the runtime checks. Context and memory aren't an afterthought either: they're a managed, versioned resource per agent, so you can see exactly what reaches a prompt and rip out what you want it to forget. That's the actual trick here — the model isn't deterministic, it never will be, but the workflow around it can be. Same shape, same rules, every run — that's what "graph engineering" means to us, as opposed to praying over a prompt. And spend is per-agent, so you can finally catch the setup that's "cheap" only because nobody's looking closely. One Go binary. Apache-2.0. Zero telemetry, zero phone-home — we don't want your data, we just want the harness to not suck. Tell us if this clicks or needs more unpacking. We can take it.

Cogitorium

Agent workflows with a model per agent — and a leash on each

Agent workflows with a model per agent — and a leash on each