Most enterprise AI conversations still start with the model. The faster way to build something that actually ships is to start with the workflow. The model is a runtime detail — the orchestration is the system.
Agents are functions; workflows are programs
A single agent invocation looks a lot like a function call: input, reasoning, output. Useful, but bounded. The interesting work — booking, reconciling, escalating, deciding — happens when many of these calls compose into a workflow that survives retries, partial failures, and concurrent execution.
In enterprise systems, "compose" is doing heavy lifting. It means: persistence between steps, audit traces for every decision, idempotency, rollback paths, and a way to reason about cost-per-run. That is not model engineering. It is workflow engineering with a model in the loop.
The four properties we design for
- Determinism at the workflow level — non-determinism is allowed inside an agent step, but the workflow re-runs must converge.
- Replayability — any past run can be re-executed against a snapshot to validate or debug.
- Observable cost surface — every step has a measurable cost (latency, tokens, downstream API calls).
- Failure isolation — one bad step does not poison the workflow; it gets quarantined and routed.
Why we built our own orchestration runtime
We tried existing frameworks. They are excellent for hackathon-scale assembly and fragile at production scale — too many of them collapse the distinction between "the agent" and "the workflow". The orchestrator should be calmer than the agents it runs. It should be the boring, audited piece. The agents bring the intelligence; the runtime brings the discipline.
“The agent is the interesting part. The runtime should be the boring part. That asymmetry is what makes this safe to run against real enterprise traffic.”
What this unlocks
Once orchestration is solid, model substitution becomes a configuration change. New agents plug into existing workflows without rewriting the workflow. Cost optimization moves to where it belongs — at the workflow planner, not in the prompt. And the people running the business get one place to see what the system is doing.
That is the foundation under everything we ship.