AI agent orchestration — many agents, one dependable run
One agent handles a task. A real process needs several, in the right order, with something sensible happening when one of them fails at three in the morning. That is orchestration, and most of it is not about the AI at all.
Why one agent is rarely enough
A typical accounts-payable process reads the document, checks the arithmetic, asks whether this supplier has ever been seen before, gets a person to approve anything unusual, writes the result to a ledger and mails a summary. Each of those is a different job with different tools. Handing all of it to a single prompt is how automations become unpredictable.
What an orchestrator has to handle
- Order and dependencies — a step waits for the steps it needs, and only those.
- Parallel groups — independent steps run at the same time instead of queueing politely.
- Timeouts and retries per step — a hanging tool is a failed step, not a frozen process.
- Conditions — branch on what the previous step produced.
- Human approval — the run pauses and notifies; nothing continues until a person says so.
- Sub-workflows — a whole process reused as one step inside another.
- Resume — restart a failed run from the step that broke, not from the beginning.
In AIK all of this is assembled in the browser on a card. There is no orchestration language to learn and no configuration file to keep in sync with it.
Each agent gets its own brain and its own tools
A step is an agent, and every agent has its own model and toolset: a cheap local model for sorting, a strong cloud model for the one hard judgement, a script for the deterministic part. Set the model per agent, or let the platform choose by how hard the task is — with the fallback order under your control rather than fixed in the code.
Unattended runs that behave like attended ones
A workflow can run on a schedule, from a webhook, or when another workflow succeeds. The unattended path is built from exactly the same definition as the button: same steps, same model, same keys, same approval gates, same marking of the output. Automation that quietly behaves differently at 3am is worse than none.
What every run leaves behind
- A step-by-step log — what ran, in what order, what it produced, what it cost in tokens and in money.
- A failure notice — to the address that workflow specifies, with a link straight to the run.
- A signed action certificate — a receipt for one run that anyone can verify on a public page without an account.
- An agent passport — a signed record of an agent’s identity and track record, embeddable on your own site as a trust badge.
Run it where the data is
Self-host the whole thing — orchestrator, agents, tool servers, database — or use the hosted version. Tool servers run as separate processes, so a broken tool takes out its own step and nothing else.