← Back to Workshop Hub USE CASE DEEP DIVE · MULTI-AGENT CASE STUDY

When Agents Work Together

One healthcare-governance problem — keeping internal policies in step with new MOH circulars — solved end-to-end by a team of cooperating agents. A worked case study in the orchestration pattern, grounded in real documents.

📖 ~15 min read ⚡ Interactive 📜 Policy & Circular Compliance 🧭 Roadmap concept

📜 A Problem That Doesn't Fit in One Box

In Workflow Patterns (D1) you met orchestration — a coordinator that decides the plan as it runs and pauses at human gates. This case study puts that pattern to work on one real healthcare-governance problem: keeping Synapxe's internal policies and SOPs in step with new MOH circulars. It's a job no single-purpose agent handles well, because it spans retrieval, interpretation, comparison, drafting and review — each a different skill.

🎯
This is a recognition exercise, not a build. The goal is to see how a coordinator plus specialists cooperate — so when the DNA / IT team proposes a multi-agent design, you can read it, question it, and set the governance gates.
The Scenario

📨 A New Circular Lands

A new MOH circular arrives — say, an update to data-handling or infection-control requirements. Somewhere in Synapxe's policy library, an unknown number of internal policies and SOPs may now be out of date. Today a compliance officer has to work out which documents are affected, read the circular against each one, decide what must change, draft the updates, and get them signed off — largely by hand, and largely in sequence.

The manual reality today
The officer opens the circular, searches the library for anything that looks related, reads each policy line-by-line against the new rules, keeps a private list of gaps, drafts changes in a document, then emails the right owners for review. Nothing is parallel, the trail lives in one person's inbox, and a missed policy only surfaces at the next audit.

🤖 One Agent vs a Team of Agents

The Guideline Assistant you built in Foundations is a single grounded agent — brilliant at one focused task, like answering a question from one policy. The compliance cycle is bigger: it needs several kinds of expertise working together and reporting to a coordinator. That's a multi-agent system.

SINGLE AGENT · ONE FOCUSED TASK MULTI-AGENT · A COORDINATED TEAM 🌊 Grounded agent Answer from one policy · today on AgentSea 🎼 Coordinator 📚 Retrieve 🔬 Compare ✍️ Draft Specialists cooperate · a roadmap concept
🌊

What a single agent is great at

One clear job, grounded in your documents, with a human verifying. This is where you start — and where most value lands first.

👥

What a team adds

Different specialists (retrieve, interpret, compare, draft, review) run in coordination under one coordinator, so a cross-cutting job gets done in one pass.

🧑‍⚖️

What stays the same

A human still owns the decision. More agents means more discipline about where a person approves — not less.

Where AgentSea is today AgentSea today runs a single grounded agent with skills like Document Analysis and Document Builder. The multi-agent coordination in this case study — a coordinator directing specialists — is a roadmap direction, a concept subject to platform readiness. You can design it now, and build the pieces today as separate grounded agents that a person stitches together.
The Cast

👥 Seven Roles, One Goal

A multi-agent system mirrors how a good compliance team already works: a coordinator who owns the outcome, a set of specialists who each do one thing well, and a reviewer who checks the whole before it goes to a human. Each agent has its own instructions, its own grounding documents, and its own tools — just like a real person has their own remit. Click any card to see what it does, what it reads, and what it hands on.

🎼
Supervisor

Compliance Coordinator

Reads the circular, decides who does what, resolves conflicts, assembles the brief.

📚
Specialist

Retrieval Agent

Finds the internal policies & SOPs the circular might touch — grounded in the library.

⚖️
Specialist

Requirements Agent

Extracts the concrete obligations from the circular — what must change, and by when.

🔬
Specialist

Gap & Impact Agent

Compares each policy to the new rules, finds the gaps, flags who's affected.

✍️
Specialist

Drafting Agent

Drafts the revised policy text and a plain-language change notice — with citations.

Reviewer

QA & Consolidation

Cross-checks the specialists, builds the traceability matrix, prepares the sign-off pack.

🧑‍⚖️
Human gate

Compliance Officer

Not an agent — the person who reviews, decides, and signs off. Always in the loop.

🧠 Why Split the Work Like This?

You could try to make one giant agent do everything. In practice, splitting into specialists is what makes the system trustworthy and maintainable — the same reasons you'd split the work across a real team.

🎭

Focused instructions

Each agent's system prompt does one job, so it's shorter, sharper, and less likely to drift than one prompt trying to do five things.

📚

Local grounding

The Retrieval agent knows the library; the Requirements agent knows how to read a circular. Each is grounded in exactly the documents it needs.

🔍

Testable in isolation

You can check the Gap agent on its own — feed it a policy and a rule, see if it finds the gap. Easier to trust, easier to fix.

🧑‍⚖️

Clear accountability

When the coordinator finds a conflict it can't resolve, it escalates to the human — rather than one opaque model guessing.

The Orchestration

🎼 Watch the Team Coordinate

This is the whole system in one picture. The coordinator reads the circular, spawns the specialists it needs, gathers their work, and routes it through a reviewer to a human approval gate before anything is finalised. Click any agent to see its moment, or press auto-play to watch a circular flow from the officer's desk to a signed-off compliance brief.

SPECIALISTS · SPAWNED AS NEEDED REQUESTOR 🧑‍⚕️ Compliance Officer COORDINATOR 🎼 Compliance Coordinator 📚 Retrieve policies from the library ⚖️ Interpret circular extract obligations 🔬 Gap & impact compare vs. new rules ✍️ Draft updates redline + change notice QA & consolidate traceability matrix 🧑‍⚖️ Human approval gate DELIVERABLE 📋 Compliance brief

🧭 What Makes This "Orchestration"

The coordinator doesn't follow a fixed script — it decides the plan from what the circular contains. A one-line data-retention tweak might touch two SOPs; a new infection-control standard might touch twenty. Same team, different plan each time.

Plan emerges at runtime Specialists spawned as needed Human gate on high-impact change Every step cited & logged
🔗
Straight from D1: this is the orchestration pattern — a coordinator plus specialists with a human approval gate. If you explored the orchestration tab there, this is that same shape, now zoomed all the way in on one problem.
End-to-End

🎬 Follow One Circular Through

Here's the same flow told as a story — a single MOH circular travelling from the officer's desk to a signed-off brief. Notice where the work runs together, where the coordinator decides, and where a human stays in control.

📨Circular in 🎼Plan 👥Specialists Consolidate 🧑‍⚖️Sign-off
1
Officer submitsThe compliance officer drops the new MOH circular into the coordinator with one instruction: "Check this against our policy library and prepare the updates."
2
Coordinator plansThe coordinator reads the circular, works out which policy areas are in scope, and decides which specialists to spawn — the plan depends on what the circular actually says.
3
Specialists work togetherRetrieval pulls the candidate policies and SOPs; Requirements extracts the new obligations; Gap & Impact compares each policy to the rules and flags who's affected; Drafting prepares redlines and a change notice — each grounded in the real documents, each citing its source.
4
Coordinator resolves conflictsIf two specialists disagree — a gap is flagged but not drafted, or a requirement maps to no existing policy — the coordinator reconciles it, or escalates the question rather than guessing.
5
QA consolidatesThe reviewer cross-checks every specialist output and builds a traceability matrix — each circular clause linked to the affected policy, the proposed change, and an owner.
6
Human approvesHigh-impact changes stop at the approval gate. The officer reviews the brief, checks the citations, adjusts, and signs off. Nothing is finalised without this step.

🗂️ The Output — A Traceability Matrix

The deliverable isn't a wall of text — it's a table a human can audit at a glance. Every new requirement is tied to a document, a change, and an owner, with the source cited. This is illustrative of the shape of the output, not real Synapxe policy data.

Circular clauseAffected policy / SOPProposed changeOwner · status
Data-retention period updatedRecords Management SOPRevise retention schedule; cite new clauseRecords · draft
New breach-notification windowData Incident Response PolicyShorten notification timeline; add stepInfoSec · draft
Access-review frequencyAccess Control PolicyGap — no current clause; new section neededIT · flag
Staff awareness requirementTraining & Awareness SOPAdd mandatory refresher on the new ruleHR · draft

⚖️ The Shape of the Change

The point of the team isn't a magic number — it's a change in shape. Manual work is sequential and lives in one inbox; the multi-agent flow runs the reviews together and leaves a complete, cited trail for a human to approve.

🐌 Manual, today

  • Search the library by hand — easy to miss a policy
  • Read each document against the circular one at a time
  • Gaps tracked in a private list; trail lives in email
  • A missed policy surfaces at the next audit

⚡ With a coordinated team

  • Retrieval sweeps the whole library, grounded
  • Reviews run together, not one-by-one
  • Every clause traced to a policy, change and owner
  • Full cited trail ready for human sign-off
🧑‍⚖️
The human still decides. The team compresses the legwork — finding, comparing, drafting, tracing. Whether a policy actually changes is a judgement call that stays with the compliance officer.
Why It Wins Here

🏆 When a Team Beats a Soloist

Policy & circular compliance is a textbook fit for multi-agent, for the same reasons a good manager splits work across a team rather than asking one person to do it all.

🎭

Domain expertise stays local

The Requirements agent reads circulars; the Retrieval agent knows the library. Each is grounded in exactly its own documents — no one agent has to be good at everything.

Parallel where it can be

Once the coordinator has the candidate policies, the per-policy comparisons and drafts don't wait in a queue — they proceed together.

🔄

Adaptive planning

The coordinator sizes the job to the circular. A small tweak spawns two sub-tasks; a major standard spawns many. Nothing is over- or under-done.

🔍

Traceable by design

Every clause is tied to a source and an owner. That audit trail is exactly what a governance function needs — and what a single black-box answer can't give.

⚠️ The Real Trade-offs

More agents means more power and more to manage. These are the costs to weigh before reaching for a multi-agent design — the same ones the DNA / IT team will raise.

🔍

Lower transparency

Seven agents conversing is harder to follow than one. You need agent traces and logs of every step — plan for observability from day one.

⏱️

Latency stacks up

Even run together, each specialist call takes time. Fine for a compliance cycle measured in hours — not for real-time chat.

🎯

More prompts to maintain

Several system prompts and tool sets to version. A change in one specialist can ripple — treat prompts like code.

🧑‍⚖️

Human-in-the-loop is mandatory

For compliance, accountability sits with a person. The team drafts and traces; a human approves. That gate is not optional.

So — Reach for a Team, or Not?

✅ A team fits when…

  • The job genuinely spans several kinds of expertise
  • The plan can't be fully scripted up front
  • Sub-tasks can run in parallel
  • A traceable, auditable output matters
  • There's a clear point for a human to approve

❌ Stick with one agent when…

  • One focused task solves the problem (start here)
  • The steps are fixed — chain them instead
  • You can't yet define the human approval gates
  • Full transparency of every step is a hard requirement
  • Speed matters more than breadth
🎯
Rule of thumb: don't reach for a team by default. Start with the simplest thing that works — one grounded agent — and only add coordination when a problem genuinely needs several specialists. Many first designs are one good agent in disguise.
Today vs Roadmap

🧭 How to Use This — Now and Next

You've seen the full team. The honest picture: AgentSea today runs a single grounded agent, so you don't wire up seven cooperating agents this afternoon. What you can do is design the system now, and deliver the value in pieces with today's tools — then adopt coordination as the platform matures.

✅ You can do this today

Build the pieces as single grounded agents

  • Build the Requirements reader as one agent — paste a circular, get the obligations
  • Build a Gap checker — one policy + the rules, does it comply?
  • Use Document Builder to turn the findings into a branded brief
  • A human plays "coordinator" — running the pieces and stitching the result
Under the hood — how DNA / IT would build the full team A coordinator model that plans and calls specialist sub-agents, grounded in a shared knowledge base, with approval gates. On AWS: Bedrock AgentCore Runtime for the managed agent runtime and memory, Strands Agents for the sub-agent SDK, and a Bedrock Knowledge Base for grounding. On AgentSea today, you build the individual grounded agents; the automatic coordination is the roadmap direction.

🧩 Quick Recap — The Whole Pattern

PieceIts one jobWhy it's separate
🎼 CoordinatorPlan, dispatch, resolve, assembleOwns the outcome and decides the plan at runtime
👥 SpecialistsRetrieve · interpret · compare · draftEach grounded in its own docs, testable on its own
✅ ReviewerCross-check, build the matrixCatches conflicts before a human sees them
🧑‍⚖️ Human gateReview, decide, sign offAccountability stays with a person — always

📐 Your Next Step — Label Your Canvas

When you fill in your Agent Design Canvas, ask whether your use case is really one agent or a small team. If it's a team, name the coordinator, list the specialists, and — most importantly — mark where a human must approve. That labelled design is your brief to the DNA / IT team.

🧠
Leader's takeaway
Multi-agent isn't about replacing the team — it's a model of the team. The coordinator does the routing, the specialists do the legwork, and you decide where human judgement and sign-off belong. Start with one grounded agent; graduate to a coordinated team only when the problem earns it.
🔭
Where this sits: D1 gave you the four patterns; this case study zoomed in on orchestration with a full worked example. Next, take it to your own use case on the Agent Design Canvas.