Orchestration — from isolated agents to one nightly chain
The first agents ran in isolation. Each was right, together they contradicted themselves. The fix was not a smarter model — it was a chain with order and hand-off.
Once the source was available, the first agents emerged. One interpreted contract terms. One reviewed invoice lines. One cross-checked attendance. Each agent, alone, was surprisingly good.
Together they were a mess.
What went wrong
The contract agent said a customer was entitled to indexation on 1 January. The invoicing agent saw no indexation in the invoices — and reported a shortfall. Meanwhile the attendance agent saw fewer hours and adjusted the other way. Three contradictory signals on one case, each technically correct within its own scope.
There was no missing intelligence. There was missing orchestration.
What worked
We stopped shipping isolated agents and built one nightly chain with explicit order:
- Ingestion — fresh restore, semantic layer, integrity checks.
- Contract understanding — which terms apply for this period?
- Expected invoicing — what should the invoice be?
- Actual invoicing — what did the invoice become?
- Reconciliation — where do 3 and 4 diverge, and how many euro is it worth?
- Explanation — which contract clause or event explains the deviation?
- Report — ordered by euro impact, with an evidence link to source.
Every step got an explicit input, explicit output, and the promise that the next step would not start before the previous was verified. No agent got more authority than its step.
The real lesson
Agents only work once you accept that an agent is not a colleague. A colleague knows to wait for someone else. An agent just does what you ask, regardless of what the neighbour concluded. If the chain does not enforce order, the system produces contradictory truths — neatly justified.
THE LESSON The breakthrough was not a smarter model. It was accepting that orchestration is a discipline: explicit steps, explicit hand-off, no agent above its station.