← All Labs
MentatIn DevFramework

Agents with careers.

Every agent framework treats an agent as a static config — same capabilities on day 1 and day 1,000. Mentat gives agents careers: they earn trust, get promoted, and spawn and mentor juniors — all on a graph that stores the system's intelligence instead of bolting memory on afterward. It's the framework behind Ziggy.

See how it works ↓
FIG. 05 AGENTS WITH CAREERS SPAWNS JR JR MID SR ↑ LEAD PRINCIPAL

The reframe

TENET 02 / 04

Keep the relationships — then generalize what held up.

The graph is the operating system. Mentat is what we extracted from Ziggy once the coordination bet proved out in production. All four tenets →

The problem

Every agentic framework solves the same starter problem — get an LLM to use tools and follow instructions — and they all work, for simple agents. But the moment you want agents that improve over time, grow into new responsibilities, and teach each other, they hit the same wall: the data model was never built for it, and bolting it on creates the kind of architectural debt that makes senior engineers quit.

The solution

Mentat makes the graph the operating system. Agent profiles, conversations, tools, feedback, and every routing decision live as nodes and edges — so “which agent is best for this task?” is one scored query, not infrastructure you maintain. On that foundation, agents have careers, the system generates its own training data, and guardrails work at both the prompt and runtime level.

How it works

What it does differently.

1
The graph is the OS

Agents, conversations, tools, and feedback all live in one graph. Picking the right agent is a single scored query — not infrastructure you build and babysit.

2
Agents have careers

Junior to Principal. Agents earn trust from real outcomes, get promoted automatically, and spawn juniors that inherit their specialization and hard-won lessons.

3
It trains itself

Every message emits training data as a side effect — prompt snapshots, context versions, routing decisions — exportable to fine-tune the system on its own history.

The build story

"Every framework hits the same wall — agents that can't grow. We built Mentat because we hit it ourselves."

Mentat was extracted from Ziggy — a production system handling real agent orchestration, not a weekend hackathon. It's the long bet under the rest of Labs: Gristle proved graph-first structure holds up, Ziggy put it into production coordinating agents, and Mentat generalizes it into a framework where the architecture, not the operator, decides how the system grows. TypeScript-first, graph all the way down.

Where this is headed.

Mentat is the framework behind how we think agents should be built. Read the white paper for the full argument, or talk to us if this is the kind of problem you're working on too.