LLM & Agentic AI

Multi-Agent Systems Without the Hype

See when a coordinator and scoped specialists earn their cost, and when one bounded agent should write the result.

Reference architecture - illustrative, not a client deployment

Explain delegation, bounded coordination and when one agent is sufficient.

Complexity
Intermediate
Drawing
Supplied
  • Agentic AI
  • Multi-agent
  • Orchestration
  • Evaluation

Overview

This reference architecture is illustrative. It is not a record of a client deployment.

A coordinator should delegate only when one context cannot hold the work. Anthropic's orchestrator-workers workflow uses a central model to split a task, hand the pieces to worker models and synthesize the results. Their research system gives each subagent an objective, an output format, tool guidance and task boundaries, in its own context window. There is no shared memory.

A concrete example

A product manager asks whether a vendor's published uptime claim is supported. Research returns findings and sources for that question only. Analysis states what those findings do and do not establish. Writing later receives the approved evidence, not the raw search trail. A deterministic review checks that the draft names its sources and does not fill a gap with an unsourced figure. If research exceeds its budget, the result says the evidence is incomplete.

Common misconception

"If several agents agree, the answer is verified." Agreement is not proof. A checker can test structure and whether citations resolve. Judging the substance may still need a person. Anthropic reported that early orchestrators spawned dozens of subagents for simple queries, and that multi-agent runs in their data used about 15 times as many tokens as chat.

Diagram & Walkthrough

Choose a declared scenario and step, or select a component by pointer, touch or keyboard. Playback starts only when you press Play and pauses when this panel leaves the screen.

100%
Multi-Agent Systems Without the Hype Delegation, bounded coordination and when one agent is sufficient. A coordinator plans inside runtime budgets and hands each specialist an explicit scoped task. Review checks structure before a synthesis or an honest partial result. Specialists do not share memory, and agreement is not proof. Multi-Agent Systems Without the Hype Delegation, bounded coordination, and when one agent is sufficient. COORDINATION SCOPED SPECIALISTS REVIEW AND DELIVERY Goal Bounded task plan Scoped research task Scoped analysis task Findings + sources Analysis + assumptions Approved evidence Draft synthesis Reviewed result Simple task, one agent Budget or specialist timeout HUMAN Task request MODEL Coordinator DETERMINISTIC Runtime budgets + handoffs steps · time · tokens MODEL Research specialist MODEL Analysis specialist MODEL Writing specialist own scoped context only DETERMINISTIC Evidence + quality review Checks structure. A person judges substance. No shared memory. EXTERNAL Synthesis or partial result Legend Data Control Feedback Alternate Timeout Agreement is not proof Explicit handoffs only. Coordination adds steps, latency and tokens.

Main walkthrough, step 1 of 6: Assess task complexity. idle.

Main walkthrough

The primary path through this reference architecture.

Active parts: Task request, Coordinator, Runtime budgets + handoffs

Text walkthrough

This vertical reference remains readable without the interactive controls.

  1. 01Assess task complexity

    The coordinator reads the goal and proposes a bounded plan. The runtime, not the model, will enforce the limits.

    Active parts: Task request, Coordinator, Runtime budgets + handoffs

  2. 02Delegate with isolated contexts

    Research and analysis each receive a scoped task. Neither inherits the other's context or a shared scratchpad.

    Active parts: Runtime budgets + handoffs, Research specialist, Analysis specialist

  3. 03Collect evidence

    Findings, sources, analysis and assumptions arrive as separate handoffs for structural review.

    Active parts: Research specialist, Analysis specialist, Evidence + quality review

  4. 04Compose a synthesis

    Writing receives the approved evidence and drafts from that package alone.

    Active parts: Evidence + quality review, Writing specialist

  5. 05Review the result

    The draft returns for a structural check. Passing the check does not prove the claim is true.

    Active parts: Writing specialist, Evidence + quality review

  6. 06Deliver with limitations

    The reviewed synthesis is delivered with its remaining limitations visible.

    Active parts: Evidence + quality review, Synthesis or partial result

Connections

  • GoalTask request to Coordinatordata
  • Bounded task planCoordinator to Runtime budgets + handoffscontrol
  • Scoped research taskRuntime budgets + handoffs to Research specialistcontrol
  • Scoped analysis taskRuntime budgets + handoffs to Analysis specialistcontrol
  • Findings + sourcesResearch specialist to Evidence + quality reviewdata
  • Analysis + assumptionsAnalysis specialist to Evidence + quality reviewdata
  • Approved evidenceEvidence + quality review to Writing specialistdata
  • Draft synthesisWriting specialist to Evidence + quality reviewdata
  • Reviewed resultEvidence + quality review to Synthesis or partial resultdata
  • Simple task, one agentRuntime budgets + handoffs to Writing specialistcontrol
  • Budget or specialist timeoutRuntime budgets + handoffs to Synthesis or partial resultcontrol

Components

Task requesthuman

A person states the goal. The request does not itself choose specialists or authorize unlimited model calls.

Responsibilities

  • State the goal and any constraints the person already knows

Inputs

  • User goal

Outputs

  • Task request

Risks

  • Ambiguous success criteria
  • Hidden constraints that never reach the plan
Coordinatormodel

Proposes a bounded plan and whether specialist work would change the result. It does not share a hidden memory with the specialists.

Responsibilities

  • Decide if one agent is enough
  • Name specialist tasks, boundaries and expected outputs

Inputs

  • Task request

Outputs

  • Bounded task plan

Risks

  • Over-delegation of a simple task
  • Vague instructions that cause duplicate work
Runtime budgets + handoffsdeterministic

Enforces step, time and token limits and records each handoff. The model cannot raise its own budget.

Responsibilities

  • Cap steps, time and tokens
  • Dispatch only the scoped context for each handoff

Inputs

  • Bounded task plan

Outputs

  • Scoped specialist tasks
  • Single-agent task
  • Timeout signal

Risks

  • Limits set so high that coordination cost is invisible
  • A dropped handoff that looks like a completed result
Research specialistmodel

Collects findings and sources for one scoped question. It does not see the analysis or writing context.

Responsibilities

  • Search within the assigned question
  • Return findings with source references

Inputs

  • Scoped research task

Outputs

  • Findings and sources

Risks

  • Duplicated searches from an underspecified task
  • Sources that do not answer the assigned question
Analysis specialistmodel

States what the assigned material supports and which assumptions remain open. It does not treat agreement with another specialist as verification.

Responsibilities

  • Separate supported claims from assumptions
  • Name gaps instead of filling them

Inputs

  • Scoped analysis task

Outputs

  • Analysis and assumptions

Risks

  • Overclaiming beyond the supplied evidence
  • Smuggling unsourced facts into the assumptions
Writing specialistmodel

Drafts from the context it is given: approved evidence, or one simple task when specialists add no value.

Responsibilities

  • Write only from the provided context
  • Keep unsupported points out of the draft

Inputs

  • Approved evidence
  • Or a simple single-agent task

Outputs

  • Draft synthesis

Risks

  • Smoothing over a missing source
  • Using context that was never approved
Evidence + quality reviewdeterministic

Checks structure: required sections, resolved citations and explicit gaps. It does not prove the substance is true.

Responsibilities

  • Check structure and citation resolution
  • Block a draft that hides a missing specialist result

Inputs

  • Findings and sources
  • Analysis and assumptions
  • Draft synthesis

Outputs

  • Approved evidence
  • Reviewed result

Risks

  • A passing structure around an unsupported claim
  • Treating the checker as a substitute for human review
Synthesis or partial resultexternal

Returns the reviewed synthesis, or a partial result that says which evidence is missing after a budget or specialist timeout.

Responsibilities

  • Deliver the reviewed result
  • Label incomplete evidence as incomplete

Inputs

  • Reviewed result
  • Or a timeout from the runtime budget

Outputs

  • User-visible synthesis or partial result

Risks

  • Presenting a partial result as a finished answer
  • Omitting the limitation

Scenario reference

Single-agent alternative

Use one bounded agent when specialization adds no clear value.

  1. 01Plan a simple task

    The same goal and budget path decides that specialists are unnecessary.

    Active parts: Task request, Coordinator, Runtime budgets + handoffs

  2. 02Hand the task to one agent

    Writing receives the simple task directly. Research and analysis stay idle.

    Active parts: Runtime budgets + handoffs, Writing specialist

  3. 03Review and deliver the draft

    The single draft still passes structural review before it is delivered.

    Active parts: Writing specialist, Evidence + quality review, Synthesis or partial result

Specialist timeout

Surface incomplete evidence rather than silently inventing the missing result.

  1. 01Delegate the research task

    Research is started with a scoped task and a finite budget.

    Active parts: Runtime budgets + handoffs, Research specialist

  2. 02Return the incomplete result

    When the budget or the specialist times out, the answer states the gap. It does not invent the missing findings.

    Active parts: Runtime budgets + handoffs, Synthesis or partial result

Design Decisions

  • Pass a scoped task and only the context each specialist needs.
  • Keep budgets, handoffs and timeouts in deterministic code. The coordinator proposes a plan; it does not raise its own limits.
  • Send a simple rewrite to one writing agent when extra specialists would not change the result.
  • On timeout, name the missing output. Do not invent it or present the partial result as complete.

Risks & Trade-offs

Vague delegation makes specialists duplicate searches or miss the question. Extra model calls add latency and token use, so Azure's orchestration guidance says to stay at the lowest complexity that still works. A structural check can pass a fluent draft whose sources do not support the claim. Specialists that must share every intermediate decision are a poor fit for this split.

When to Use / Not Use

Use scoped specialists for independent investigations, separate tool permissions, or more evidence than one context can hold. Do not use them for a fixed lookup, a single rewrite, or steps that all depend on one running conversation. Start with one bounded agent, and add coordination only when the simpler path fails.

Sources

Related