Task decomposition and routing
A complex request is broken into verifiable steps and handed to the right agent, instead of being dropped whole on one model that improvises.
AI agent orchestration from prototype to production
An agent that answers well in a demo is not a production system. Here is what it takes to hold up on real processes, real data and a controlled budget.
Why a single agent is not enough
The demo nearly always works. You wire a model to a few tools, give it access to the data, and the answers look excellent. Then the system meets the real process: dirty data, edge cases, ambiguous requests, a write operation nobody authorized and a token bill nobody can explain.
The difference between a prototype and a production agentic system is not the model: it is everything around it. An LLM is never wired straight to a database. Between the request and the effect sit task decomposition, strict schema validation, permissions and a record of what happened.
Orchestration means exactly this: deciding which agent does what, with which tools, within which limits and under whose human oversight.
The pieces of orchestration
The components that separate an experiment from a system the business can rely on
Guardrails and schema validation
Every output entering company systems goes through strict validation. If it does not match the schema it does not pass: it is not patched downstream in the hope that it works out.
Human-in-the-loop
High-risk and write operations stop at a human approval. Automation reaches as far as the cost of a mistake stays acceptable, and no further.
Observability and cost per operation
Every run leaves an immutable trace: context read, data extracted, steps taken and tokens spent. That is what makes cost predictable and errors reconstructable.
Provider abstraction
Business logic is never tied to one model vendor. An abstraction layer lets the model be swapped the way any other component is swapped.
Data access and RAG
Agents read company data through controlled retrieval, with permissions and per-tenant segregation. The quality of the sources decides the quality of the result.
How we put it on the ground
It starts from the process, not from the model. First you look at where the work jams today, which data actually exists and in what state, which decisions can be automated and which have to stay human. The orchestration is designed from there, the model is chosen after, and the system is built around the real constraints.
Rollout is incremental: one process at a time, measured before and after. An agentic project that cannot show what it improved is a project that gets switched off at the first budget cut.
A real case: agents running audit interviews
We built a multi-tenant platform where contextual AI agents run audit and compliance interviews on ISO 27001 and GDPR, collecting structured evidence in place of static questionnaires. The measured numbers: 60% more data collected, 88% accuracy and four hours saved per audit.
What is the difference between an agentic workflow and classic automation?
Classic automation runs steps decided in advance: change the shape of the input and it breaks. An agentic workflow decides how to reach the goal within the limits it was given, and can handle input that varies and is unstructured. In exchange it needs far more control: without guardrails and traceability, flexibility turns into unpredictability.
How do you stop an agent writing to the wrong systems?
An LLM is never wired straight to a database, even if that can look like the simple route. Agents act through explicit tools, with limited permissions and strict data-schema validation; write operations and high-risk ones go through a human approval before they take effect. Every step produces an audit log that stays accessible and checkable.
How do you keep token costs under control?
Through observability at the step level: every operation carries what it cost. That makes it possible to use different models for different tasks, cut pointless hops and set per-process limits, instead of discovering consumption at the end of the month.
Is our data used to train the models?
No. We select enterprise cloud services that exclude it contractually, in line with GDPR and SOC 2. Where data is particularly sensitive, a private AI infrastructure is considered, on dedicated machines or even on-premise.
What happens when a new model comes out?
Nothing dramatic, the architecture is provider-agnostic. Business logic sits outside the model, so replacing it is a component swap with a verification phase, not a rewrite of the platform.
Insights
All articles Working on agents, governance and AI architectures
An agentic system reaches production one step at a time.
You set the goals, weigh the routes and build only what holds in production. If an agent is not worth it today, we say so.
- You won't talk to a salesperson
- NDA before we get into the details
- The code stays yours