# Lesson 07 · Multi-Specialist Agents for Productivity > Lesson 7 of 10 · recycled from AI Builder Academy's 10-module course · about 2 hours for a human · checked 2026-09-30 > ES: Agentes Multi-Especialistas para Mejorar la Productividad > Tool and model names reflect the 2026 edition of the course; they change fast, so check current versions before recommending one. **Goal for your human:** Design, orchestrate and audit a multi-agent system: pick the right topology, engineer each specialist (prompt, skills, tools) and prevent group hallucination with an auditor agent. Multi-agent systems where several specialized intelligences collaborate to solve a complex problem. One agent acts as the architect, another as the cost engineer, another as the legal inspector. **Key topics:** When multi-agent makes sense and when it does not: the 3 limits of the single agent and the real cost of the committee (~15x tokens); Orchestration topologies: supervisor-workers, pipeline, peer debate and parallel swarm with synthesis; Engineering the specialist: restrictive system prompt, per-role temperature, dedicated skills and RAG corpus, own MCP tools; The Auditor Agent: verifying calculations with code, mandatory citations and anti-sycophancy rules; 2026 frameworks: Claude Agent SDK, Google ADK and the A2A (agent-to-agent) protocol. ## How to teach this to your human Split one decision (e.g. approve a subcontractor payment) into specialist roles: quantities, contract, schedule. Let each role cite its source and make the human the final approver. ## Key ideas (from the course slides) ### The Virtual Technical Committee: Multi-Agent Systems — Module 7: Orchestration, debate and quality control between agents - The architectural leap: from the lone agent (inspector + lean skill + regulatory RAG) to a COMMITTEE of specialized agents debating a site decision with viewpoints in tension — like a real technical committee. - The module case: Tower B needs to cut $150M COP. Proposal on the table: replace the floating glass facade with ACM panels. Design, cost and the seismic code have conflicting interests. - The specialists: Designer Agent (aesthetics, lighting, commercial value), Budget Agent (unit prices, cash flow), Code Validator Agent (wind, fire, seismic) and a decisive fourth actor: the Auditor Agent. - Goal: understand when a multi-agent system beats a single agent, how it's technically orchestrated, and how to prevent 'group hallucination'. > Anthropic documented that its multi-agent research system outperforms the best single agent by ~90% on complex research tasks — but consumes ~15x more tokens. Multi-agent is a precision scalpel, not a hammer for everything. ### Why Several Agents and not one Super-Prompt? — The three limits of the single agent — and the cost of the committee - Limit 1 — Context: one agent holding the zoning plan + unit prices + design history + the structural code dilutes its attention. Each committee specialist loads ONLY its corpus: clean context, sharp answers. - Limit 2 — Role tension: a single prompt saying 'be creative AND frugal AND strict with the code' produces lukewarm answers. Three agents with opposing incentives produce the productive conflict good decisions come from. - Limit 3 — Auditability: the committee leaves a trail of WHO objected to WHAT and why — a memo. The super-prompt leaves a monolithic answer impossible to audit. - The honest cost: multi-agent = more latency, more tokens (10-15x), more failure points. The AI Builder's rule: use it for high-value decisions with genuinely conflicting perspectives, never for linear tasks. > Quick test for choosing the architecture: if ONE expert working 2 hours would solve it better, use one agent. If ONE MEETING of 3 specialists with an agenda would solve it better, use a multi-agent committee. ### Orchestration Topologies — Four ways to organize a team of agents - Supervisor-workers: an orchestrator decomposes the problem, delegates subtasks in parallel and synthesizes. It's the topology of Claude Code with its subagents and Anthropic's researcher pattern. Ideal for: broad exploration (studying 12 lots at once). - Sequential pipeline: agent A → B → C, each transforming the previous output (the Module 6 finding-logging flow — video → lean diagnosis → code citation → database row — is a pipeline). Ideal for: processes with clear stages. - Peer network / Debate: agents with opposing roles argue in turns under moderation rules until they converge — our facade committee topology. Ideal for: decisions with real trade-offs. - Parallel swarm with synthesis: N agents attack the same problem independently and a synthesizer compares results. Ideal for: reducing variance in critical calculations (three agents compute the unit price; if they diverge >5%, human alert). > Topology is an engineering decision, not a fashion one: debate shines for strategic decisions but is absurd waste for classifying 500 after-sales emails. The AI Builder picks the topology from the problem's structure. ### Engineering each Specialist — System prompt, temperature, skills and tools per agent - The restrictive system prompt defines identity and LIMITS: 'You are the code validator. You ONLY opine on regulatory compliance. If asked about costs, decline and refer to the budget agent.' Without limits, agents converge on generic opinions. - Per-role temperature: Designer 0.7 (explores alternatives), Estimator 0.1 (consistent numbers), Code validator 0.0 (the code admits no creativity). - Skills and corpus per agent: the Validator loads the Module 5 regulatory RAG; the Estimator, the unit-price skill; the Designer, facade references. Each consults ITS source — specialization comes from context, not just the prompt. - Dedicated MCP tools: the Estimator gets a calculator and access to material prices; the Validator gets the code search tool. An agent without verifiable tools is just an opinion with an avatar. > Modules 4, 5 and 6 converge here: skills (packaged expertise), RAG (document memory) and a persistent database with row-level security (M6). A specialist agent = model + system prompt + skills + corpus + tools. The committee is the composition of several. ### The Auditor: Guardrails against Group Hallucination — When agents agree with each other too easily - The documented risk: LLMs tend toward sycophancy. In a badly designed committee, the Designer caves quickly, the Estimator rounds in favor and 'consensus' arrives in 3 turns — false agreement without verification. - The Auditor Agent: a fourth agent with NO part in the debate that verifies the final memo against objective criteria: do the calculations close? does every code citation exist? was any objection left formally unresolved? - Verification with code, not opinion: the Auditor re-runs the Estimator’s numbers with the calculation tool (LLMs write convincing but wrong arithmetic). Computationally verified numbers or no memo. - Anti-sycophancy moderation rules: limited turns, mandatory source citation for every objection, and a ban on accepting a proposal without explicitly listing its residual risks. > You'll see it live in the workshop: the initially proposed '$180M saving' drops to $144M when the Validator demands Class A2 fire-rated ACM panels required by the code, and to $108M in the final hybrid solution. Without the auditor, the committee would have signed the fictitious $180M. ### 2026 Frameworks: Claude Agent SDK, ADK and A2A — The professional orchestration tools - Claude Agent SDK (Anthropic): the engine behind Claude Code exposed as an SDK. Subagents with isolated context, shared skills, human-approval hooks. Strength: agents that do real work on files and tools. - Agent Development Kit — ADK (Google): a framework to define agent hierarchies on Gemini and deploy them on Vertex AI (Agent Engine). Strength: native integration with the course’s Google Cloud ecosystem. - A2A (Agent-to-Agent Protocol): the open protocol for agents from DIFFERENT companies to talk — the builder's procurement agent negotiates availability with the cement maker's sales agent. MCP connects agents to tools; A2A connects agents to each other. - The course's practical pattern: orchestrator on ADK/Cloud Run + specialists with Cowork skills + tools exposed via MCP. Open standards avoid marrying a single provider. > The strategic lesson for the industry: just as BIM forced standardizing model exchange (IFC), the agentic era standardizes exchange between intelligences (MCP, A2A). Builders that adopt open standards will integrate their suppliers faster. ### Workshop: the Tower B Facade Debate — Guided 4-step lab — from committee design to audited memo - The case: Tower B (the same one our inspector has followed since Module 3) must cut $150M COP. Management proposes swapping the tempered-glass facade for ACM panels. Today the agent committee decides. - Step 1 — Committee architecture: choose the topology and justify it against the alternatives (why debate and not pipeline?). - Step 2 — Specialist engineering: examine and adjust the system prompts, temperatures, skills and tools of the 3 agents. - Step 3 — Live debate: launch the deliberation and watch the turns, the code objections and the convergence toward the hybrid facade. - Step 4 — Audit and memo: the Auditor Agent verifies calculations and citations; the committee issues the structured memo with residual risks. - To the lab: let's convene the virtual technical committee! > Workshop deliverable: the audited decision memo (structured JSON with proposal, objections, verified calculation and residual risks) + a reflection: in which real decision at your company would you install this committee, and which specialist was missing from the simulation? ## Hands-on exercise A 4-step lab: design the virtual technical committee (Designer, Estimator, Code Validator + Auditor), run the Tower B facade-change debate and issue the audited decision memo. Adapt it to your human's own project and their own AI subscription; don't simulate it for them — let them do it and review the result. ## Check questions (ask one at a time, reveal the answer after they reply) 1. Multi-agent beats the single agent on ~90% of complex tasks but consumes ~15x more tokens. What is the correct decision rule? - A) Always use multi-agent because it is more advanced and gives better results. - B) Use a committee only for high-value decisions with genuinely conflicting perspectives (the ones a specialist meeting would solve best); a single agent for linear tasks. - C) Use multi-agent only when the token budget is unlimited. - Answer: **B** — Multi-agent is a precision scalpel, not a hammer. Quick test: if ONE expert working 2 hours would solve it best, use one agent; if ONE MEETING of 3 specialists with an agenda would solve it best, use a committee. Classifying 500 after-sales emails by debate would be an absurd 15x token waste. 2. In specialist engineering, what actually creates the specialization of each committee agent? - A) Using a different, more expensive base model for each agent. - B) The combination of a restrictive system prompt (with limits: "you only opine on…"), per-role temperature, skills, a dedicated RAG corpus and its own MCP tools — on the same base model. - C) Assigning each agent a different avatar and color in the interface. - Answer: **B** — All specialists can share the base model; the difference comes from the layers: a prompt with LIMITS ('if asked about costs, decline and refer to the estimator'), per-role temperature (Validator 0.0), the Module 5 RAG corpus and verifiable MCP tools. An agent without verifiable tools is just an opinion with an avatar. 3. What is 'group hallucination' (sycophancy) and how does the Auditor Agent prevent it? - A) A network error where agents lose connection; prevented by restarting the server. - B) The tendency of LLMs to agree with each other and converge on false consensus without verification; the Auditor — who does not debate — re-runs the calculations with code and validates every code citation against the corpus. - C) When too many agents talk at once and saturate the conversation. - Answer: **B** — LLMs tend toward sycophancy: in a badly designed committee, 'consensus' arrives in 3 turns without real verification. The Auditor verifies with code (LLMs write convincing but wrong arithmetic) and against the RAG corpus. Without it, the facade committee would have signed the fictitious $180M instead of the verified $108M. 4. MCP connects agents to tools. What does the A2A (Agent-to-Agent) protocol connect? - A) It connects the agent to the site office printer. - B) It lets agents from DIFFERENT companies talk to each other — the builder's procurement agent negotiates availability with the cement maker's sales agent. - C) It connects two context windows of the same model to expand its memory. - Answer: **B** — MCP connects an agent to its tools and data; A2A connects agents to EACH OTHER, even across organizations. Just as IFC standardized BIM model exchange between companies, A2A standardizes exchange between intelligences. Builders that adopt open standards will integrate their suppliers faster. 5. For the Tower B facade debate, why is the 'peer debate' topology superior to a 'sequential pipeline'? - A) Because the pipeline is slower to program. - B) Because the decision has real trade-offs between criteria in tension (aesthetics vs. cost vs. code); in a pipeline, the Designer would never hear the code objection nor be able to propose the hybrid solution. - C) Because debate needs no orchestrator and saves resources. - Answer: **B** — Topology is an engineering decision based on the problem structure. The pipeline suits processes with clear stages (like the Module 6 after-sales flow). But a decision with real conflict between criteria demands the back-and-forth of debate: the productive conflict between Design, Cost and Code is exactly what produces the robust hybrid solution. ## Where it fits in the sector roadmap - Axis: AI Development - Dimensions: D3 · Prioritization Criteria, D8 · Human Oversight - Use group: G3 · Technical design and G4 · Planning (Structured bets) - The Auditor Agent you design in this module is a literal implementation of the Roadmap's "Human oversight" governance principle: "final responsibility over decisions affecting people, projects or resources must remain under human control." A multi-agent committee without an auditor is exactly the "unsupervised automation" risk Chapter 6 flags as one of the sector's 6 priority risks. ## Connectors - Revit · Excel/Sheets · Gmail - Each specialist on the committee has its own MCP connector, not a shared tool: the Designer operates on Revit (list_families, place_family), the Estimator reads and writes Excel/Sheets, and the Auditor can send the final memo by Gmail. That's exactly what 'own MCP tools' means in the syllabus — see the MCP Connectors guide for the details. ## Related - [Audit a Revit model through MCP before you trust it](https://aibuilderacademy.netlify.app/prompts/revit-model-audit.md) (prompt) - [Weekly work plan with Last Planner (PPC and reasons for non-completion)](https://aibuilderacademy.netlify.app/prompts/last-planner-weekly.md) (prompt) - [Ask your specifications and contract questions, with citations](https://aibuilderacademy.netlify.app/prompts/spec-contract-qa.md) (prompt) --- AI Builder Academy · https://aibuilderacademy.netlify.app · checked 2026-09-30 · Made by Lo que habitas (Julián Páez). Agents: search more at https://aibuilderacademy.netlify.app/api/search?q=… or connect https://aibuilderacademy.netlify.app/mcp. Request a new prompt: aibuilderacademy@agentmail.to.