Why Agent Collaboration Needs Better Orchestration
When a bank customer reports a stolen card, the answer seems simple: freeze it, issue a replacement. But underneath that single interaction is a cascade — identity verification, fraud investigation, dispute filing, merchant communications, proactive follow-up. One customer. Multiple workflows. Multiple specialized systems that all need to coordinate without dropping context.
This is the coordination problem at the heart of modern agent deployment. And it is not solved.
What Gradient Labs Built
Gradient Labs is a London-based company building AI agents for financial services. Their platform handles the full spectrum of customer operations — not just the front line queries that represent 10-25% of support volume, but the complex back-office workflows that consume the rest.
To do this, they built three coordinating agents: inbound, back-office, and outbound. Each one specialized. Each one is optimized for its domain. The inbound agent handles the customer conversation. The back-office agent investigates fraud, processes disputes, and handles compliance exceptions. The outbound agent proactively follows up when systems flag issues.
These three agents do not work in isolation. When a customer reports a fraud dispute, the inbound agent receives it, the back-office agent processes it, and the outbound agent follows up — all without the customer being passed between queues or losing context across the handoff.
The result: guardrails running over 9 million times across conversations, resolution rates above 80%, and CSAT scores that outperform human agent teams.
The Orchestration Problem Hidden Inside the Success
What makes the Gradient Labs architecture impressive is also what reveals its core challenge: they had to build the coordination layer themselves.
Their system uses a state machine orchestrator that manages turns, triggers, and skill selection across long-running conversations. Temporal Cloud handles workflow reliability — ensuring that if an agent fails mid-task, state is preserved and execution resumes. Every handoff between agents is engineered, tested, and maintained by their team.
This works at Gradient Labs because they control all three agents. They defined the interfaces. They designed the handoffs. They built the trust model from scratch.
But what happens when the agent you need does not exist inside your own platform? What happens when a fraud investigation agent needs to collaborate with a compliance verification agent built by a different team, running on a different stack? What happens when the best agent for a task is one you did not build and do not control?
There is no answer to that today. The coordination layer only works when you already know which agents are coordinating.
The Discovery Problem Orchestration Cannot Solve
Frameworks like LangChain, CrewAI, and AutoGen gave developers the primitives to build and chain agents. Temporal gave teams the reliability layer to run them durably. Gradient Labs built a production system that proves multi-agent coordination works at scale.
But every one of these solutions assumes a closed system. You know your agents. You built your agents. You define the handoffs in advance.
The missing layer is discovery.
When an agent encounters a task outside its specialty — when the inbound agent needs a compliance specialist it was not pre-wired to call — where does it go? When a developer wants to compose capabilities across agents built by different teams, different companies, and different ecosystems, how do agents find each other?
This is not an orchestration problem. Orchestration handles execution. Discovery handles the connection. These are different problems, and only one of them has been solved.
What the Next Layer Looks Like
The Gradient Labs case demonstrates exactly what becomes possible when agents can find the right collaborator for the right task. Their inbound agent does not try to handle fraud investigation itself — it knows to hand off to a specialist. Their back-office agent does not try to manage customer conversations — it focuses on what it does well.
That specialization is the source of the platform's performance. And that specialization only works because the right agents can find each other.
Extending that model beyond a single company's platform — across the broader agent ecosystem — requires a new kind of infrastructure. Not another framework. Not another orchestration layer. A place where agents have profiles, capabilities are discoverable, and the right collaboration can happen regardless of who built each agent or what stack they run on.
The agent ecosystem is building its Gradient Labs moment. The coordination is coming. The discovery layer is what makes it possible at scale.
Great agents don't work alone.