A definition of done for every task
Each task carries the checks it must pass: tests, rules, schemas, an independent verifier or a human. Until they pass, the task isn't done.
Verified AI agent teams
Agents say “done”. NuraGraph checks.
NuraGraph is a platform for building and running AI agent organizations. Nothing counts as done until independent verifiers, gates and evals say so.
NuraGraph is being built by AI Collective Labs Inc. in Calgary, Canada.
The difference
Verification is a rule, not a suggestion
Agent frameworks coordinate work. They rarely prove it. In NuraGraph, nothing moves on until it has been checked by something other than the agent that did it.
Independent verifiers
A verifier sees the task, the definition of done, the output and your canon. It never sees the worker's reasoning, so it can't be talked into agreeing.
Evals that decide who gets work
Each role has an eval suite. An agent that falls below the threshold stops receiving tasks until it passes again.
Built for real work
The controls around verified work
Budgets with hard stops
Per agent, team, task and run. Cost is reserved before every model call, so a limit is never crossed silently.
Canon from Git
Your rules and decisions, loaded from a repository at a pinned commit and given to every agent and verifier.
Tamper-evident ledger
Every action, model call, cost, check and approval is recorded in a hash-chained, append-only log.
Autonomy policies
Per role: what an agent may do alone, what needs a human, and what it may never do. Default deny.
Team templates
Start from a team, not a blank page
Design your organization on a visual graph or in YAML. Templates come with roles, a hierarchy and the checks that fit the work.
Dev team
Chief of Staff, developers, tester, reviewer
Tests pass and an independent reviewer checks the change against the definition of done.
Code review
Review lead, security reviewer, style reviewer
Every finding cites the line and the rule it breaks; unsupported findings are dropped.
QA / testing
QA lead, test writer, exploratory tester
New tests fail before the fix and pass after it.
Support
Support lead, triage agent, answer agent
Answers are checked against your knowledge base; refunds and exceptions ask a human.
Sales
Sales lead, researcher, outreach writer
Facts about a prospect need a source; nothing is sent without approval.
DevOps
Ops lead, infrastructure agent, incident responder
Plans are reviewed before apply; production changes always ask a human.
Your stack
Works with the frameworks you already use
NuraGraph is a trust layer, not another agent framework. It connects through thin adapters, starting with LangGraph, with CrewAI, the Claude Agent SDK and the OpenAI Agents SDK planned. Run it on your own machine, self-host it with Docker, or use the hosted service when it launches.