Architecture Discussion: Structuring a Multi-Agent Procurement Negotiation Simulator

[TathastuNaranje]

Feature Area

Core functionality

Is your feature request related to a an existing bug? Please link it here.

Hi CrewAI team and community,

I'm a Market Intelligence Consultant exploring the application of agentic workflows in enterprise procurement. I'm currently designing a "Negotiation Rehearsal" system and would love to get your thoughts on the optimal architecture using CrewAI.

Describe the solution you'd like

Sourcing managers need to roleplay high-stakes negotiations before meeting with strategic suppliers. The system needs to simulate the supplier's sales team, armed with their public financials, recent earnings call transcripts, and known market pressures.

Proposed Agent Structure:

The Supplier Persona Agent: Instructed to act as the supplier's Account Executive. It holds the context of the supplier's current market position (e.g., "We need to defend margins this quarter").

The Market Analyst Agent: A background agent that feeds real-time commodity pricing or macro trends into the simulation, altering the supplier's leverage dynamically.

The Coach Agent: Observes the interaction between the human sourcing manager and the Supplier Persona, providing real-time feedback on negotiation tactics.

My Questions:

•
What is the best pattern in CrewAI to have the "Market Analyst" asynchronously update the context of the "Supplier Persona" during an active chat session?

•
Are there examples of using the HierarchicalProcess where the "Coach" acts as a manager, evaluating the human's input before passing it to the "Supplier Persona"?

Any insights or pointers to similar implementations would be incredibly valuable. Thanks for building such a powerful framework!

Tathastu Naranje

Describe alternatives you've considered

No response

Additional context

No response

Willingness to Contribute

Yes, I'd be happy to submit a pull request

[xg-gh-25]

Market Analyst feeding Supplier Persona: CrewAI's CrewOutput.pydantic gives you shareable state. Have the Analyst emit structured context (e.g., MarketConditions(commodity_price_trend: float, supplier_leverage: str)), then inject it into the Supplier's system prompt via context=[analyst_task]. Update frequency: run the Analyst on a conditional trigger (e.g., every 10 human turns) to refresh leverage dynamically.

Coach-as-Manager with HierarchicalProcess: This is harder than it sounds — the manager_llm in Process.hierarchical orchestrates task delegation, not human-input interception. You'd need to wrap the human's message in a custom tool the Coach reviews before forwarding. Cleaner pattern: make the Coach a post-interaction observer — it reads the full conversation history (via context) after each exchange and surfaces feedback as a separate output, not inline approval.

The single-agent alternative (proven in production): Multi-agent coordination has real costs — state synchronization, feedback loops, and debugging "who said what." We run a role-switching single agent for similar simulations: one LLM that explicitly declares current_role: supplier | coach | analyst in its output schema, with tools that toggle roles mid-conversation. The memory layer holds the full context; roles are just lenses. Simpler state, one conversation thread, and you can still get adversarial negotiation tactics (the "supplier" role gets adversarial instructions). If your primary goal is training humans (not multi-agent research), consider this first.

Production lesson from our procurement pilots: The Supplier Persona needs an exit strategy — without it, the LLM negotiates forever because it has no "walk away" threshold. We hard-code a reservation_price in the Supplier's context; if the human crosses it, the agent explicitly ends negotiation. Procurement managers found this more realistic than infinite flexibility.

Solid architecture question — the fact you're asking about feedback loops and async state (not just "how do I chain agents") shows you're thinking about the real problem.

---

Multi-agent vs role-switching patterns from SwarmAI. Discussion: T-MvS