AI Agentic vs Single-LLM: When Orchestrators Beat Prompts
Three quarters of Romanian mid-market teams that tried to automate a compliance workflow last year did it with a single-LLM prompt chain. Most quietly rolled back. The reason is not model quality — it is topology.
AI Agentic vs Single-LLM: When Orchestrators Beat Prompts
Three quarters of Romanian mid-market teams that tried to automate a compliance workflow last year did it with a single-LLM prompt chain. Most quietly rolled back. The reason is not model quality — it is topology.
A single-LLM approach hands the whole process to one context window: read the invoice, check ANAF rules, decide the fiscal code, write the response. It works on toy problems. On a real RO business process — say a DAC7 marketplace reporting run, or an e-Factura reconciliation cycle governed by OPANAF 1.581/2025 — the context blows up, tool calls collide, and every retry rewrites earlier decisions.
Agentic architecture splits the same job across role-specific workers coordinated by an orchestrator. Each agent owns one sub-goal, one toolset, one memory scope. The orchestrator arbitrates and keeps the audit trail. That is the shape that survives an ANAF inspection.
When single-LLM is still the right answer
If the process fits in one prompt, uses one API, and needs no branching — keep it single-LLM. Classification of a support ticket, extraction of fields from a fixed template, translation of a clause. Adding an orchestrator here only adds latency and a second failure surface. The European Commission’s AI adoption guidance is explicit that added architectural complexity without a decision-tree gain is waste.
The moment the process has more than three branches, two tool families, or an approval step, single-LLM starts to lose. Token cost climbs, reasoning drifts, and there is no clean per-decision record — a problem for GDPR Art. 22 automated-decision documentation and for the transparency obligations under the EU AI Act.
What a working orchestrator looks like
orchestrator:
process: e-factura-reconciliation
agents:
- name: invoice_parser
tools: [xml_reader, cui_lookup]
- name: anaf_validator
tools: [spv_api, opanaf_ruleset_2025]
- name: reconciler
tools: [ledger_query, mismatch_writer]
policy:
max_retries_per_agent: 2
escalate_to_human_after: 2
audit_log: /var/log/orchestrator/2026-09.jsonlEach agent runs with its own tool budget and its own logs. When an ANAF ruleset updates, only the anaf_validator prompt changes — the rest of the graph is untouched. NIST’s AI risk management framework recommends exactly this kind of decoupling for testability, and ENISA’s 2024 threat landscape treats agent isolation as a control against prompt-injection blast radius.
The CAI Technology position
We build orchestrators when the process has legal reviewability. If a Romanian tax inspector, an ANSPDCP auditor, or an internal risk committee needs to trace why an outcome happened, one agent per decision boundary beats one prompt per everything. Model choice comes second. Topology comes first. Our practice on multi-agent design patterns connects directly to how we handle retrieval pipelines under audit — same discipline, different surface. If your 2027 automation programme is still one giant prompt, talk to us before you scale it.