AI agents: simplify architecture while preserving controls
How many agents does it take to handle a customer case? For a Belgian or French company, business acceptance tests should determine the answer. Anthropic’s October 22, 2026 guide offers a starting point for comparing architectures, without assuming the outcome on an organisation’s own data.
1. The official starting point
Anthropic proposes evaluating whether some subagents can be folded into a main loop, loading procedures and tools on demand. The guide retains a role for subagents in verification, parallel work and permission separation. It recommends gradual migration measured through end-to-end evaluations.
2. What changes for Belgian and French companies
Our recommendation: an SME should start with a narrow process, such as drafting a reply from authorised tickets and documents. Mid-market and large enterprises should compare architectures on the same cases, with the same permissions and model. Public administrations should include access refusals and justification requirements in acceptance testing. IT and business teams should jointly agree on what counts as a correctly completed task.
3. Underside analysis: separate procedure from authorisation
We recommend treating a procedure loaded by an agent as work instructions, never as permission. For RAG, the document service should filter every search according to the user’s rights. For Odoo, tools should distinguish reading, proposing a change and performing an approved write. Combining reasoning does not justify combining all privileges in one service account. Plan logs that connect the request, documents accessed, tool calls and human approval.
4. Acceptance testing before migration
Build a representative set of cases: a straightforward answer, a missing document, conflicting sources, a document injection attempt and a forbidden write. Measure business success, correct references, expected refusals, latency and cost per completed case. Replay cases several times, then change one boundary between agents. Retain a specialist architecture if simplification weakens a control or an answer. Plan a rollback with versioned procedures, model and tools.
5. Keep sovereignty in the decision
Our view: a simpler architecture should also be evaluated for supplier dependency. For local, private or cloud AI, document separately where requests, documents, memory and logs travel. Test trace export and access revocation before expanding the pilot. The decision should concern an operable, reversible service with an identified owner, rather than the advertised number of agents.
Operational recommendation: select a read-only workflow, set acceptance thresholds before testing and compare both architectures. Allow business writes only after validating permissions, traces and rollback.
Scope a governed AI project