Configuration over code paths
Every workflow expressed as a code path is a workflow the customer cannot change. Describing composition as configuration moves the ceiling from what has been anticipated to what can be assembled.
Multi-agent orchestration
Configurable coordination of several specialised agents inside a governed workflow, instead of routing every task through one general-purpose agent.
- Organisation
- Salesforce / MuleSoft
- Period
- 2025 – Present
- Area
- Agentic AI infrastructure
Context
Enterprise workflows tend to decompose naturally: one agent understands a domain, another has the tool access, a third performs a check. A single general-purpose agent handling all of it is harder to reason about and harder to constrain.
The problem
Composing agents raises problems a single agent never has. Tasks have to be routed to the right specialist, context has to cross a boundary without leaking more than it should, failures have to be attributable to a step, and enterprise controls have to apply to the interactions between agents, not just at the outer edge of the system.
Constraints
- Agent interactions must remain governable; a chain of agents cannot become a way around the controls applied to a single one.
- Customers have to be able to configure the composition themselves. Building it per workflow does not scale.
- Failure in one step must be isolated and explicable, not surfaced as an opaque whole-workflow error.
- Context propagation has to carry identity without over-sharing it.
My contribution
Shared ownership stated as such. Nothing here claims sole authorship of a platform.
- Contributed to configurable multi-agent orchestration capabilities that let enterprises chain agents and coordinate them within structured workflows.
- Worked on how agent-to-agent interactions are configured and how tasks route between specialised agents.
- Worked on applying enterprise controls to the interactions between agents, not only at the system boundary.
Approach
Model the workflow as a graph of agents and the edges between them, so composition is described as data. The shape of a workflow becomes something a customer configures and an operator can read, not something encoded in a deployment.
Apply policy at the edges, not only at the entrance. An agent calling another agent is a governed interaction in its own right, which keeps the guarantees of a single agent intact under composition.
Propagate context explicitly. What crosses an agent boundary is a decision, which makes both identity propagation and failure attribution tractable.
Decisions that mattered
The ones with a real cost on the other side.
Isolating failure per step
Agent steps fail for very different reasons. A tool is unreachable, a model is rate-limited, a policy denied the call outright. Collapsing all of those into a single error makes the system quick to build and almost impossible to debug.
Outcome
A configurable system for composing specialised agents into structured enterprise workflows, with governance applied to the interactions between them.
What generalises
- Composition must not weaken the guarantees that hold for a single component.
- Reliable systems make failures explicit and attributable.