Case study
Chat-Based Multi-Agent Collaboration Network
Built a chat-based platform where multiple AI agents collaborate in shared conversation rooms, resolving references to people, records, and prior context automatically instead of asking the user to repeat themselves.
Overview
Beyond a single chatbot, the client needed multiple specialized agents that could work together in the same conversation, each understanding what "that contact," "the deal from earlier," or "the sheet I mentioned" actually referred to.
The Challenge
Chat-based AI usually treats every message in isolation, or forces the user to re-specify exact IDs for every record it acts on. That breaks down fast once more than one agent is involved, or once a conversation runs long enough that the referenced context isn't in the last message.
The business impact:
- Users had to repeat exact names, IDs, or details every time an agent needed to act on something mentioned earlier.
- Coordinating more than one agent in a single conversation had no shared sense of identity or room context.
- A vague reference like "that folder" or "the spreadsheet" had no reliable way to resolve to the actual record.
Approach
How do multiple agents collaborate in one conversation without stepping on each other?
Single-agent chat doesn't scale to real work that spans specialties. We built an agent network on an A2A-style (agent-to-agent) protocol — every agent has a resolvable identity, can be addressed directly by any other agent, and can hand off, relay, or route work through the network rather than just replying in a single thread.
How do you turn "that contact" or "the sheet from earlier" into a real record?
Vague references are how people actually talk, but APIs need exact IDs. We built entity and ID resolution across the integration layer — folders, sheets, contacts, presentations, chat threads — so an agent resolves a natural reference to the record it actually points to instead of asking the user to look up an ID.
How do you keep a human in control when agents are proposing actions to each other?
Letting agents hand work to each other freely is powerful and risky in the same breath. We built approval previews into the collaboration flow, so on sensitive actions, a human reviews and signs off before anything executes, not after.
Tech Stack
Outcome
Start a conversation
Let's build yours.
Book a 20-minute discovery call — we'll tell you honestly whether AI is the right tool for what you're trying to solve.