AI Multi-Agent Orchestration Platform
A production agentic system that decomposes complex tasks, dispatches specialised agents, maintains working memory, and routes tool calls through a typed execution graph.
Reduced task failure rate from 34% to under 4% on multi-step document processing workflows. Mean time to diagnose a production failure dropped from 45 minutes to under 8 minutes.
Context bleed between agent handoffs was corrupting task state on multi-step workflows.
When orchestrating chains of specialised agents, task context accumulated noise across handoffs — previous agent reasoning, partial results, and tool call artifacts bled into downstream prompts. This caused non-deterministic failures impossible to reproduce in isolation: a task that succeeded in testing would silently degrade in production because the context it received was always slightly different.
A typed execution graph with explicit memory boundaries at every agent transition.
Rather than passing raw context strings between agents, I built a typed execution graph where each agent receives only its designated input schema — nothing more. Working memory is managed separately from execution context: a shared memory store handles cross-agent state, while each agent's prompt is constructed programmatically from typed fields. Alternatives included prompt chaining (too fragile) and a single monolithic agent (too constrained by context limits).
More orchestration overhead per task, but non-deterministic failures became traceable failures.
The typed graph adds latency at each node boundary — around 40-60ms of serialisation and schema validation per transition. In exchange, every agent handoff is inspectable: the exact input and output of each agent is logged with a correlation ID. A failure that would have been "the agent behaved strangely" is now "agent X received malformed input from agent Y at step 3, here is the diff." The overhead is worth it in production.