[FEATURE] Support structured orchestration as a first-class agent behaviour

Preflight Checklist

• [x] I have searched existing requests and this feature hasn't been requested yet
• [x] This is a single feature request (not multiple features)

Problem Statement

Claude Code's Agent tool allows an agent to spawn subagents, but there is no structured platform support for task orchestration — decomposing a complex task into sub-tasks, dispatching multiple agents in parallel or sequence based on dependencies, and synthesising their outputs into a single coherent result. This pattern currently works only when the orchestrating agent spontaneously decides to do it in-context, which is unreliable and undiscoverable.

Proposed Solution

Introduce a structured orchestration primitive — either as a declarable schema or a first-class response type — that allows an agent to express:

Decomposition: how a task is broken into named sub-tasks with ownership
Parallelism: which sub-tasks can run concurrently
Dependencies: which sub-tasks require other sub-tasks' outputs before they can begin
Collation: which agent synthesises the final output, and how
Conceptual form:

orchestrate:
parallel:

• agent: bertram-gilfoyle

task: audit infrastructure and security posture

• agent: jared-dunn

task: build financial model and data room
then:
sequential:

• agent: richard-hendricks

task: document algorithm (requires: gilfoyle's audit)

• agent: dinesh-chugtai

task: clean up API and developer docs
collate: erlich-bachman
The exact form is open — the point is that decomposition, dispatch, dependency, and collation should be declarable, not left entirely to in-context model improvisation.

Alternative Solutions

Ad hoc Agent tool calls: An agent can already spawn multiple subagents manually. Works when the model decides to do it, but not structural, not reliable, and produces inconsistent results across invocations.

Chained delegation: Each agent delegates to the next sequentially. Loses parallelism and produces no synthesis — the final agent's output is the result, with no collating voice.

Orchestration via hook: A UserPromptSubmit hook decomposes the task before the agent sees it. Same problem as delegation via hook — routing is invisible, the orchestrating agent never makes the decision, and the synthesised output loses the orchestrator's voice and judgment.

Priority

Medium - Would be very helpful

Feature Category

Other

Use Case Example

A user asks Erlich Bachman to prepare Pied Piper for a Series B. This is a complex task that spans multiple domains:

Infrastructure and security posture → Gilfoyle
Financial model and data room → Jared
Algorithm documentation and technical differentiators → Richard
API cleanup and developer docs → Dinesh
Investor narrative, framing, and pitch → Erlich himself
With structured orchestration:

Erlich decomposes the task, identifies what he owns (narrative, pitch) and what belongs to his team
He dispatches Gilfoyle, Jared, Richard, and Dinesh — in parallel where possible
Once all outputs are collected, Erlich synthesises them into a single investor-ready document in his voice
Each section carries the character and expertise of the agent who produced it
Without this feature, either Erlich does everything himself (wrong), or the orchestration happens ad hoc with no guarantee of decomposition, parallelism, or synthesis.

Additional Context

Orchestration is distinct from delegation (Issue 64483) in an important way: an orchestrating agent is not passing off a task because it's out of their domain. It is actively directing multiple agents, retaining ownership of the final output, and synthesising the results. The orchestrating agent still does the work within their own expertise — they only distribute the parts that belong to others.

These two features are complementary. Delegation handles the single-hop case (wrong agent, find the right one). Orchestration handles the multi-hop case (right agent, but the task is too large for one agent alone). Both are needed for a fully capable multi-agent team.

Multi agent setup where the requested feature would be useful: https://github.com/cliffano/tech-team-agents