3.0 KiB
3.0 KiB
status
| status |
|---|
| proposed |
One Orchestration engine above the agents
Every request for work is one of several kinds: an operating model, a solution architecture, both, or a question about work already delivered. Each kind belongs to an agent, and each agent is the Execution engine and one Workflow. The KAB platform does not route between agents; its workflows call one agent per step. Without something above the agents, every caller has to know which agent a request is for and talk to it directly, and "both" cannot be done at all.
We decide that every request comes through one Orchestration engine, a service of its own in this repository, beside the Execution engine package. It is an A2A agent: callers, the KAB platform among them, register and call it as they would any agent.
- It decides which agent the work belongs to, and produces nothing itself. No Deliverable, no Phase, no Approval of its own.
- It knows its agents by configuration, a list of the kinds of work and the address of the agent that does each. Adding an agent is a line of configuration, not code.
- It dispatches the request unchanged. The message and its metadata go to the chosen agent as they came, but for the field that named the work.
- It relays. Every question and Approval the agent pauses for is passed to the caller, and every reply back, unchanged. It holds which agent's task each of its own tasks is waiting on.
- It does not guess. In order: what the caller named settles it (
work); with only one agent configured, that agent; otherwise it starts nothing and says how to name the work. Reading a request by Jev, with its confidence thresholds, is step F3 and is added to these rules, not in place of them. - One Customer case across agents. The same slug goes to each agent a request reaches.
Proposed 2026-10-20.
Considered Options
- Each caller calls the agents directly: rejected. Every caller would carry the routing, and "both" (architecture, then the operating model from its files) would be built once per caller.
- Route in the KAB platform's workflows: rejected. The platform calls one agent per step and cannot relay an agent's questions to a person; and decision 6 of the implementation plan keeps routing out of the platform.
- A router inside the Execution engine: rejected. The engine holds one Execution of one Workflow (ADR-0001); choosing between agents is above any one Execution.
Consequences
- The Orchestration engine is a fourth container beside the agents and is the only one registered with the KAB platform.
- A caller that pauses for a person sees every agent's questions and Approvals through it. A caller that cannot answer, such as a platform workflow step, sends
pre_approvedas it would to the agent. - Its own tasks map to agents' tasks in memory. A restart loses tasks in flight; the agents keep their checkpoints.
- While an agent works, it reports that it is still working at an interval, so a caller that treats a silent stream as dead does not retry and start a second Execution.