Agent orchestration
MadMax as an agent orchestration platform: AI agents with tools, memory and knowledge, routed by switches and loops, joined by merges, gated by human approvals — every hand-off a declared shape, every run verified before it goes live.
"Agent orchestration" means more than one AI agent doing real work, with something deciding which agent runs, what it gets, and what happens with the result. MadMax is that something — and it adds the part most agent frameworks leave to you: proof that the agents, their hand-offs and their side effects actually work before anything is live.
The orchestration primitives
Everything an orchestrator needs is a step on the canvas, built by the AI from a sentence or dragged on by hand:
| you need | the step | what it does |
|---|---|---|
| a worker | AI Agent | a prompt with structured output, tools (web search, browser, code, other workflows), optional memory and knowledge-base retrieval |
| a router | If, Switch, Classify | send each input down the branch its content deserves — a rule, a case list, or an AI label |
| fan-out and join | Loop, Loop until, Merge | run a body per item or until a condition holds, then fan the branches back into one payload |
| a sub-agent | Call workflow | run another workflow as a function call and continue with its result — a specialist you compose |
| a human gate | Approval, Wait | stop before anything irreversible; a person decides in-app or by one-click email |
| shared state | Data stores | durable records between runs — the ledger the agents read and write |
| an entry point | Triggers | webhooks, schedules, forms, inbound email and SMS, file drops, and polling watches |
The agent that reads the invoice, the switch that routes it by amount, the approval that holds the large ones, the agent that drafts the reply, the store that remembers what was paid: that is an agent system, and it is one canvas.
What frameworks hand you, and what they don't
An orchestration library gives you the loop: call an agent, read its output, call the next. It does not tell you whether the output had the field the next agent expects, whether the tool call would have sent a real email, or whether the whole thing still works after you changed a prompt. MadMax does, because the orchestration is a workflow, and workflows here come with:
- Declared shapes on every hand-off. Each step's output is a known shape — an agent's structured output, a webhook's captured sample, an integration's documented response — so a downstream reference to a field that will not exist is caught on the canvas, not in production.
- Verification before activation. Every step is test-run on real data: agents produce real output that is judged against the plan, sends are previewed as the exact request they would make, webhooks capture a real delivery. A step is green when it is right, not when it did not crash.
- A run ledger. Every firing is a run with per-step inputs, outputs, timing and spend, so a bad decision is traceable to the agent and the input that made it.
- Spend and blast-radius guards. AI credits are metered per workspace, long-running or runaway runs are cancelled by budget, and errors route to a branch you chose.
- Humans where judgment lives. Approval is a step, not an afterthought — the orchestration pauses, a person signs, it resumes.
The builder is itself an orchestrated agent
Describing a workflow in a sentence starts an agent that plans, builds, wires, lints and verifies — and that calls its own specialists when it lacks something: a research step reads a third-party vendor's documentation for the webhook payload shape before the build guesses it, and says so when nothing documented exists. The same orchestration runs over MCP, so Claude Desktop, Cursor or your own agent can build and verify workflows through the same tools. See Build mode for the loop in detail.
What it is not (yet)
- The graph is fixed at activation. Agents choose branches, loop, and call other workflows at run time, but they do not create new steps or spawn new agents while a run is live. A dynamic sub-agent step is on the roadmap; today a specialist is a workflow you compose up front.
- Long-lived agents are sessions, not processes. Memory persists across runs per session key; an agent does not sit resident between triggers.
Where to go next
- AI workflow automation — the describe, plan, build, verify loop.
- Custom prompt — the agent step's tools, memory and knowledge.
- How verification works — what "verified" means for an agent's output.