From goal to verified result
NuraGraph runs on top of your agent framework and adds one rule: work isn't done until something other than the agent that did it says so.
- Design your organization
- Give the Chief of Staff a goal
- Agents work inside their limits
- Independent verification
- Everything is on the record
1. Design your organization
Pick a template or start blank. Arrange roles in a hierarchy on the visual graph, or edit the same configuration as YAML in Git. Each role gets its tools, policy, budget, eval suite and default checks.
2. Give the Chief of Staff a goal
The Chief of Staff agent breaks the goal into tasks, writes a definition of done for each, and delegates down the graph, only to agents whose evals are passing.
3. Agents work inside their limits
Before every action, NuraGraph checks the role's autonomy policy. Before every model call, it reserves the cost against the budgets. Anything outside policy waits for a human.
4. Independent verification
When an agent submits work, fast deterministic checks run first, then independent verifier agents that never see the worker's reasoning. A task becomes done only when every required check passes; otherwise it goes back for another attempt or to a human.
5. Everything is on the record
Each step lands in a tamper-evident ledger with its cost and evidence. The run report shows what was done and verified, what it cost, and what still needs a decision.