Multi-Agent Orchestration Fundamentals
Deliberate Academy Editorial Team
Reviewed for accuracy and professional relevance
You're 6 lessons in — don't lose your progress.
Sign up free to save where you are and earn a verified certificate when you pass.
- Design a multi-pass workflow with distinct AI roles — a drafter, a critic, and a fact-checker — using either separate conversations or dedicated agent tooling
- Write role-specific system prompts that give each pass in a multi-agent workflow a genuinely different mandate rather than a cosmetic label
- Identify agent role bleed — a critic role that rubber-stamps a draft instead of genuinely challenging it — and apply the structural fix that prevents it
- Determine when a multi-agent workflow is worth the added time and complexity, and when a single well-verified pass is the better choice
Multi-agent orchestration coordinates several distinct AI passes, each with a different role, within a single workflow — a drafter that produces a first version, a critic that challenges it, and a fact-checker that verifies its claims, run in sequence rather than as one combined request. You do not need engineering infrastructure to do this: for most professional use cases, orchestration means running separate conversations in ChatGPT, Claude, or Gemini with distinct system prompts for each role, and manually passing each stage's output into the next. Dedicated frameworks like CrewAI, AutoGen, and LangGraph automate this handoff for engineering teams building repeatable pipelines, and platform features like custom GPTs or Claude Projects can hold a role's persistent instructions — but the underlying discipline is the same whether you orchestrate it by hand or with tooling.
Why Separate Roles Catch What a Single Pass Misses
A single AI pass, even a strong one, is both the author and the only reviewer of its own work within that conversation — everything it "knows" about whether the draft is good is generated by the same reasoning process that produced the draft. Splitting a workflow into a drafter and a critic, run as genuinely separate passes with different mandates, breaks that self-referential loop. The critic pass is not reviewing its own work; it is reviewing someone else's output with a mandate specifically oriented toward finding problems, which produces meaningfully more scrutiny than asking the drafter to "double-check your own work" in the same breath that it produced the draft.
The three-role pattern that generalizes well across most professional writing and analysis tasks is drafter, critic, and fact-checker. The drafter produces a complete first version against your brief. The critic is given the draft alone — not the original brief — and asked to identify weaknesses as a skeptical reviewer would. The fact-checker is given only the specific factual claims extracted from the draft and asked to flag which ones need verification against a source. Each role sees only what it needs to do its job, which is what keeps the roles genuinely distinct.
Agent Role Bleed: When the Critic Stops Being Critical
The most common failure in a manually orchestrated multi-agent workflow is agent role bleed: the critic pass, given a system prompt that is too similar to the drafter's, or run in a conversation that still has the original drafting context visible, ends up agreeing with the draft rather than genuinely challenging it. This happens because a model's default posture toward well-organized, fluent text — including its own kind of output — is generally favorable, which is exactly the bias a critic role exists to counteract. If the critic's system prompt does not explicitly instruct it to actively look for specific categories of weakness, it tends to produce a mild, generally-approving review that adds little beyond what the drafter already believed about its own work.
A critic role that responds with general praise and only minor suggestions on most drafts has likely bled into agreement rather than genuine critique. Give the critic role a specific, adversarial mandate — "assume this draft contains at least one significant flaw; your job is to find it" — and specific categories to check, rather than an open-ended "review this and give feedback," which tends to produce exactly the mild agreement that defeats the purpose of a separate critic pass.
A Three-Pass Pipeline for Client Deliverables
Context
A content agency's client deliverables — competitive analyses and market briefs — were produced with a single AI drafting pass followed by a human editor's review. Over one quarter, editors caught an average of 4.2 factual or logical issues per deliverable during review, several of which had already gone to clients before being caught in a second look.
Action
The head of content restructured the workflow into three separate passes: a drafter conversation producing the full deliverable against the brief, a critic conversation — given only the completed draft and an adversarial mandate to find logical gaps, unsupported claims, and structural weaknesses — and a fact-checker conversation given only the extracted factual claims and asked to flag which needed a verifiable source. Each pass ran as a fully separate conversation with a distinct system prompt.
Outcome
Over the following quarter, the average number of issues human editors caught during their review dropped from 4.2 to 1.1 per deliverable, because the critic and fact-checker passes caught the majority of issues before the draft reached a human editor at all. The added time per deliverable — roughly 20 minutes for the two extra passes — was substantially less than the editor time it replaced, and no deliverable in that quarter went to a client with an issue a human editor found.
A critic pass in a multi-agent workflow consistently responds to drafts with general praise and only minor stylistic suggestions, rarely identifying a substantive weakness. What is the most likely explanation, according to this lesson?
Select one answer.
When Multi-Agent Orchestration Is Worth the Cost
A multi-pass workflow takes meaningfully longer than a single prompt and, for tasks with real complexity, is worth it — the case study above cut editor-caught issues by nearly three-quarters for roughly twenty extra minutes of process time. But not every task justifies this. For routine, low-stakes drafting — an internal status update, a straightforward reformatting task — a single well-written prompt, possibly with the verification pass from Lesson 2, is the more efficient choice. Reserve multi-agent orchestration for deliverables where a missed error has a real cost: client-facing analysis, anything entering a compliance or audit trail, and recurring high-volume outputs where a small per-item quality gain compounds into a large aggregate improvement.
In the three-role pattern, the critic is handed the draft without the original brief, and the fact-checker receives only the claims extracted from the draft. What does withholding that material achieve?
Select one answer.
Exercise
Your Task
Choose a recurring AI-assisted deliverable from your work where quality issues have occasionally slipped through to a final version. Design a three-pass workflow: write a distinct system prompt for a drafter role, a critic role with an explicitly adversarial mandate and named categories to check, and a fact-checker role that receives only the extracted factual claims. Run all three passes as separate conversations on one real example and compare the result to your usual single-pass process.
Success looks like
- The critic role's system prompt gives it a specific adversarial mandate and named categories to check, not a general 'review this' instruction
- The fact-checker role receives only the extracted claims, not the full original draft with its persuasive framing intact
- The three-pass output catches at least one issue your usual single-pass process would likely have missed
Watch out for
- Running all three roles in the same conversation thread, where the critic can still see the drafter's reasoning and full context — this reintroduces exactly the self-referential blind spot the pattern is meant to break
- Giving the critic role a vague mandate, which produces agent role bleed and mild, low-value feedback
Hint
If your critic pass keeps agreeing with the draft, try giving it an explicit count requirement — 'identify at least three specific weaknesses, even if you have to look hard for the third one' — which forces genuine scrutiny past the point where a model would otherwise stop at easy, superficial praise.
- Multi-agent orchestration runs distinct AI passes with different roles — commonly a drafter, a critic, and a fact-checker — as separate conversations or through dedicated agent tooling, breaking the self-referential loop of a single pass reviewing its own work.
- Each role should see only the information it needs for its specific job — a critic reviewing the draft alone, a fact-checker reviewing only extracted claims — which is what keeps the roles genuinely distinct rather than blended.
- Agent role bleed, where a critic defaults to agreement rather than genuine challenge, is prevented with a specific, adversarial system prompt and named categories to check, not with a general "give feedback" instruction.
- No-code orchestration through separate conversations with distinct system prompts works for most professional use cases; frameworks like CrewAI, AutoGen, and LangGraph automate the handoff for engineering teams building repeatable pipelines.
- Reserve multi-agent workflows for deliverables where a missed error carries real cost — client-facing analysis, audit-trail documentation, high-volume recurring outputs — and use a single well-verified pass for routine, low-stakes tasks.