Multi-agent orchestration and sessions

Episode Two builds on the single-agent loop by distributing work across specialised agents without losing state, authority, or evidence. By the end, you should be able to choose sensible task boundaries, decide between parallel and sequential work, construct an explicit handoff, and explain session resume and fork behaviour. Now extend that governed loop across multiple workers. Why not simply give one powerful agent every tool and the entire context? Because specialisation can reduce irrelevant context and capability, parallelise independent work, and make evaluation more targeted. Those benefits appear only when decomposition is sensible and coordination overhead does not exceed them. The Foundations guide emphasises a coordinator-and-subagent pattern.

Text description

• Frame the work: Coordinator defines the objective, relevant facts, constraints, expected output, and provenance requirements.
• Delegate bounded tasks: Selected specialists work from the context supplied rather than automatically inheriting the full conversation.
• Return structured evidence: Subagents report findings, support, uncertainty, unresolved questions, and stable source references.
• Aggregate results: Coordinator compares the returned evidence and produces the final synthesis.

The coordinator analyses the request, decomposes the task, selects subagents, passes each one adequate context, and aggregates the results. Subagents normally operate with isolated context and do not automatically inherit the coordinator's full conversation. That means delegation must include the objective, relevant facts, constraints, expected output, and provenance requirements. A vague instruction such as research this leaves the subagent to reconstruct context it cannot see. For Harbour Resolution, imagine a complex complaint involving a contract, call transcript, policy history, and transaction record. One subagent can extract contractual obligations. Another can establish the transaction chronology. A third can retrieve current policy and prior versions. A synthesis agent can compare the evidence and produce claim-source mappings.

Text description

• Parallel: independent tasks: Policy research and transaction extraction can run concurrently when neither requires the other's output.
• Sequential: dependent tasks: Final legal-risk analysis waits until the required policy and transaction results have returned.

Should the coordinator always invoke all four? No. It should select only the agents needed for the query and evidence available. A fixed full pipeline wastes time and may inject irrelevant findings. However, the coordinator should not decompose so narrowly that each worker misses cross-cutting issues. A good task boundary is independently meaningful, has clear inputs and outputs, and minimises duplicate source coverage. How do you decide between parallel and sequential delegation? Run tasks in parallel when they are independent, and sequence them when later work depends on earlier results. For example, policy research and transaction extraction may run concurrently. A final legal-risk analysis should wait until both have returned.

Text description

• Proceed with explicit gaps: Continue while clearly identifying missing evidence or unfinished work.
• Retry the failed source: Attempt the failed retrieval or task again.
• Substitute a specialist: Route the bounded task to another suitable worker.
• Escalate to a human: Transfer the issue when automated recovery is insufficient.

Parallel work also needs a strategy for partial failure. The coordinator might proceed with explicit gaps, retry a failed source, substitute another specialist, or escalate to a human. It should not silently omit a failed branch and present the synthesis as complete. What should a subagent return? A compact structured result containing findings, supporting evidence, uncertainty, unresolved questions, and stable source references. That is much better than a long narrative that forces the coordinator to rediscover structure. The coordinator should preserve source identity through every transformation. When two agents disagree, do not ask the synthesiser to average their claims. Retain both claims, compare dates and authority, retrieve more evidence where possible, and surface the unresolved conflict.

The exam often rewards explicit conflict handling over forced consensus. Could subagents communicate directly with one another? They can in some architectures, but hub-and-spoke coordination is easier to reason about, secure, and observe. The published Foundations objectives specifically centre the coordinator as the information-routing and error-handling point. Direct peer communication can create hidden dependencies, duplicated work, and provenance gaps. A professional architecture might still use a graph or peer protocol, but it must justify the extra complexity. Let us distinguish agent identity from conversation state. A session provides continuity across turns, while a subagent definition describes role, instructions, tools, and configuration. Resuming a session continues from prior state.

Text description

• Durable application state: Important case state has explicit versions and ownership outside conversation history.
• Agent session: The session provides continuity and acts as a reasoning workspace around durable state.
• Forked branch: A new traceable branch explores an alternative from an existing point and requires a merge rule.

Forking creates a new branch from an existing point so alternative work can proceed without corrupting the original path. Where is forking useful? For testing two remediation approaches, comparing prompts, or exploring a risky hypothesis while preserving the accepted branch. Each branch needs a traceable identifier and a rule for how any result is merged. Do not treat conversation history as the only system of record. Important case state should live in durable application storage with explicit versions and ownership. The agent session is a reasoning workspace around that state, not a substitute for it. Long-running research creates another problem because context accumulates as each worker returns material.

The coordinator should retain task-defining facts, evidence summaries, decisions, source mappings, and open issues. It can omit duplicated boilerplate, superseded drafts, and bulky raw results that are safely stored elsewhere. What role do hooks play around agents? Hooks intercept lifecycle events so the system can validate, log, transform, allow, deny, or trigger deterministic actions. A hook is stronger than merely asking the model to remember a rule. A pre-tool hook can block a disallowed path, inspect command risk, or attach required metadata. A post-tool hook can normalise results, record telemetry, run a formatter, or flag sensitive output. A session-start hook can prepare environment context, while a completion hook can trigger validation or notification.

Text description

• Delegation: The coordinator remains responsible, assigns bounded work, and receives the result.
• Handoff: The receiving agent or human takes primary responsibility with the case summary, evidence, attempted actions, uncertainty, and required decision.

Hooks must remain fast, observable, and predictable, because hidden hook behaviour can make agent failures difficult to diagnose. Do hooks replace authorisation inside the backend service? No. They add an enforcement layer, but every service must still validate the caller and requested operation. Defence in depth should not create a single magical interception point. Consider the difference between delegation and handoff. Delegation usually means the coordinator remains responsible and receives a bounded result. A handoff transfers primary responsibility and must pass enough state for the receiving agent or human to continue safely. Human escalation is a handoff that needs a concise case summary, evidence, actions already attempted, uncertainty, and the specific decision required.

An escalation saying the artificial intelligence system is unsure is not operationally useful. A strong multi-agent answer therefore mentions isolation, explicit context passing, purposeful decomposition, selective invocation, provenance, partial failure, and central coordination. A weak answer merely says add more agents for higher accuracy. More agents can amplify correlated mistakes, expense, and coordination complexity. Use independent perspectives where they add measurable value, not as decorative architecture. Episode Two is complete when decomposition improves focus without obscuring context, provenance, failure, or authority. For source navigation, use the Agent S D K material on subagents, sessions, user approvals, and hooks, then compare it with the multi-agent orchestration guidance.

In the next episode we will design the tools and Model Context Protocol boundaries those agents are allowed to use.