Design multi-agent AI workflows with clear boundaries, handoffs, and monitoring. Use when a complex PM task should run as parallel specialized agents instead of one linear process.

Purpose Guide product managers through designing multi agent workflows —breaking complex, repetitive PM tasks into parallel, specialized AI agents rather than linear, sequential processes or manual execution. Use this to transition from "document heavy administrator" to "systems level orchestrator" who coordinates a "living system" of AI agents, human teams, and market data interacting continuously. Key Shift: From linear project management (one task at a time) to orchestration (multiple agents working simultaneously, each with clear boundaries and handoffs). This is not about prompt writing—it's about architecting workflows where AI agents handle repetitive research, synthesis, and validation while PMs focus on strategy and decision making . Input Works best with: The workflow or recurring task you want to orchestrate — described in a sentence or two, however manual or messy it is today. Also useful: Where it breaks down now (too slow, too sequential, too dependent on you), the tools your team already uses, and whether you've worked through [context engineering advisor](../context engineering advisor/SKILL.md) first (it's the prerequisite discipline). Anything supplied with the invocation itself — text after the skill name, a pasted context dump, or an appended ARGUMENTS: line — counts as answers already given. Use it and skip whatever it covers; don't re ask. Arriving empty handed? That works too. The advisor opens by asking which PM workflow eats the most of your week, then walks the four orchestration dimensions against it. Example invocation: Design an orchestration for our weekly competitive intel: today one PM spends 6 hours scraping, summarizing, and briefing — sequentially. Key Concepts Orchestration vs. Project Management Dimension Project Management Orchestration Approach Linear oversight of schedules and human tasks Managing "living system" where AI agents, humans, and data interact continuously Task Flow Sequential (finish A, then B, then C) Parallel (A, B, C run simultaneously) PM Role Document heavy administrator Systems level leader coordinating automated systems + human judgment Focus Output (features shipped) Outcome (business results, learning velocity) Risk Management Manual tracking and mitigation Real time monitoring with agentic systems flagging gaps Critical Insight: Orchestration is not about replacing humans—it's about force multiplying human judgment by automating repetitive, time consuming tasks. The Four Dimensions of Orchestration 1. Coordination of Multi Agent Workflows Breaking complex tasks into specialized agents that run in parallel. Example: Manual (Old): PM spends 8 hours compiling competitive intel, then 4 hours synthesizing customer feedback, then 3 hours identifying roadmap gaps = 15 hours sequentially Orchestrated (New): Three agents run simultaneously: Agent A: Competitive intel (research agent) Agent B: Customer synthesis (synthesis agent) Agent C: Roadmap gap analysis (analysis agent) Total time: 8 hours (limited by slowest agent), PM reviews outputs in 2 hours = 10 hours total, 5 hours saved Key Principle: Shift from manual selection to hypothesis orchestration —agents generate hypotheses, PM validates and decides. 2. Leadership of Cross Functional AI Pods Governing diverse teams (data scientists, ML engineers, compliance, ethicists) to ensure solutions are scalable, ethical, and aligned. What it includes: Embedding diversity aware workflows Risk management (not afterthought) Ethical orchestration (ensuring AI doesn't "go rogue") Cross functional alignment (engineering, compliance, design) PM Role: Guardian of Governance—ensures AI systems reflect company values. 3. Launch Control Tower Function Real time monitoring of organizational readiness across functions using agentic systems to flag gaps before critical failures. What it monitors: Support readiness (docs, training, escalation paths) Marketing readiness (messaging, assets, GTM plan) Operations readiness (infrastructure, scaling, monitoring) Key Principle: Agentic systems act as early warning system—flag gaps before they become blockers. 4. Strategic Intent Alignment (Context Engineering Applied) Feeding AI agents the correct mix of mission, constraints, and priorities to ensure automated decisions reflect company values. Connection: This is context engineering at the orchestration layer . See context engineering advisor for foundations. What agents need: Product constraints (what we will/won't build) Strategic priorities (what matters most right now) Operational definitions (shared glossary) Evidence standards (what counts as validation) The Four AI Management Workflows (Productside Blueprint) Every PM must master these workflows to move fast while staying grounded: 1. Context Engineering ✅ (Foundation) Create AI workspace that remembers product domain, research, JTBD, personas, constraints Skill: context engineering advisor 2. Synthetic Evals 📋 (Quality Assurance) Automated validation tests for AI reasoning Generate synthetic data, run workflows against traces Eliminates 80% of hallucination risk 3. Agentic Workflows ← We're here Agents handle repetitive tasks (competitive intel, customer synthesis, roadmap gaps) PM focuses on strategy 4. Vibe Coding 📋 (Rapid Prototyping) Generate clickable prototypes from context workspace Collapse feedback loops from weeks to hours Connection: pol probe advisor (Vibe Coded PoL Probes) AI Shaped Problems (Teresa Torres) What makes a problem "AI shaped"? Previously difficult to scale due to human involvement (e.g., synthesizing 50 user interviews) Falls short with current non AI solutions (e.g., manual competitive tracking) Requires consistency at scale (e.g., risk analysis across 100 features) Key Insight: "While AI makes building easier, choosing what to build remains the primary challenge." Orchestration helps with the "building" part so PMs can focus on "choosing." The Four Big Risks (Marty Cagan, AI Era) The orchestrator manages these risks across the organization: Risk Standard Definition AI Era Layer Orchestrator's Role Value Risk Will customers pay? Does AI provide enough incremental benefit vs. cost? Validate value with PoL probes before orchestrating Usability Risk Can users figure it out? Is the UX right? (Chat often isn't!) Test workflows with real users Feasibility Risk Can we build it? Do we understand "physics of AI"? Token budget? Design within technical constraints Viability Risk Can we sell/support? Legal limits? Data privacy? Sustainable OpEx? Monitor costs, compliance, scalability Anti Patterns (What This Is NOT) Not about replacing PMs: Orchestration amplifies judgment, doesn't eliminate it Not about automating everything: Some tasks require human empathy and context Not about complexity for its own sake: Only orchestrate when it saves significant time or improves quality Not about "set it and forget it": Orchestrated workflows require monitoring and maintenance When to Use This Skill ✅ Use this when: You have repetitive PM tasks that take 5+ hours per week (competitive analysis, customer synthesis, roadmap maintenance) You're doing sequential work that could be parallelized (research, then synthesis, then analysis) You need consistency at scale (analyze 50 features for risk, synthesize 100 customer interviews) You're spending time on execution instead of strategy ❌ Don't use this when: The task is one time or infrequent (not worth orchestration overhead) Human judgment is critical at every step (empathy driven work) The task is already fast enough (don't over engineer) You haven't built context engineering foundations first (see context engineering advisor ) Application This interactive skill uses adaptive questioning to design multi agent workflows step by step. Step 0: Gather Context Agent asks: Before we design your orchestrated workflow, let's gather context: Current PM Work: What repetitive tasks take the most time each week? (competitive analysis, customer synthesis, roadmap updates, etc.) How much time do you spend on these tasks? (hours per week) Are these tasks sequential (one after another) or could they run in parallel? Desired Outcome: What would "good enough" automation look like? (80% reduction in manual work? Higher consistency?) What parts require human judgment? (final decisions, empathy, ethical considerations) Technical Context: Do you have context engineering foundations? (constraints registry, operational glossary—see context engineering advisor ) What AI tools do you currently use? (Claude Projects, ChatGPT, custom agents) You can describe briefly or paste examples. Step 1: Assess if Task is AI Shaped Agent asks: Let's determine if your task is suited for orchestration. Which statement best describes your situation? Select one: 1. "I spend 10+ hours per week on repetitive research/synthesis tasks" Examples: Compiling competitive intelligence, synthesizing customer feedback, tracking feature requests Assessment: High potential for orchestration 2. "I have a task that requires multiple sequential steps, and each step is time consuming" Examples: Research competitors → Synthesize findings → Identify gaps → Update roadmap (15 hours total, done sequentially) Assessment: Good candidate for parallelization 3. "I need consistency at scale—analyzing many similar items" Examples: Risk analysis for 50 features, synthesizing 100 user interviews, tracking 20 competitors Assessment: Perfect for orchestration (agents ensure consistency) 4. "I have a task that's mostly human judgment and empathy" Examples: Stakeholder negotiation, team coaching, ethical trade offs Assessment: Not a good candidate for orchestration (keep human driven) User response: [Selection] Agent analyzes: Based on your selection: Options 1 3: Proceed to workflow design Option 4: Recommend keeping task human driven; suggest alternative approaches (e.g., use AI for research, but keep decision making human) Step 2: Decompose Task into Agent Specific Subtasks Agent asks: Let's break your task into discrete subtasks , each handled by a specialized agent. For the task you selected, what are the distinct steps? Example Decomposition (Competitive Intelligence): Manual Process (Old): 1. Research 10 competitors' product pages (3 hours) 2. Track recent feature launches (2 hours) 3. Analyze pricing changes (2 hours) 4. Synthesize into competitive landscape doc (3 hours) 5. Identify strategic gaps (2 hours) Total: 12 hours, done sequentially Orchestrated Process (New): Agent A (Research): Scrape competitors' product pages, extract features Agent B (Launch Tracker): Monitor competitors' release notes, blog posts, social media Agent C (Pricing Analyzer): Track pricing pages, identify changes Agent D (Synthesis): Compile findings from A, B, C into structured report Agent E (Gap Analysis): Compare our roadmap vs. competitor features, flag gaps Total: Agents A C run in parallel (3 hours), then D and E run sequentially (2 hours) = 5 hours + 1 hour PM review = 6 hours total (50% time saved) Agent offers: I'll help you decompose your task. Describe your current process step by step , and I'll identify: Which steps can be agent handled Which steps can run in parallel Which steps require human judgment User response: [Describe process] Agent provides: Decomposed workflow with agent assignments. Step 3: Design Parallel vs. Sequential Flows Agent asks:

agent-orchestration-advisor