# Section 14 · Synthesis and recommendations In the end, everything comes down to one posture and five engineering moves. ## 1 Set the boundaries before the stack Define the contracts (ports) and the integration protocol first. Language and topology then become local, reversible choices, made in the right place, not initial bets. ## 2 Start simple, isolate behind contracts, add complexity on evidence Modular monolith, single orchestrator, single model port. A service is externalized, an agent is added, or a model tier is raised only when an objective criterion or a measured gain justifies it. ## 3 Keep the deterministic out of the model Classification, routing, validation, idempotence, and guardrails stay in testable, cacheable code. The model keeps its budget for what it alone does better. ## 4 Make state explicit, and memory governed An agent with no hidden state. The event journal is immutable, for audit and replay; working memory is a derived view of it, which forgets through a value score, invalidates rather than deletes, and cools off into archive. In-memory structures are externalized in order to replicate: scaling becomes a state migration, not a rewrite. ## 5 Govern, trace, and evaluate from day one A single middleware chain for logging, access, cost, and approval; a request identifier propagated everywhere; multi-provider failover and degraded mode for resilience. The same trace feeds a continuous, guarded evaluation, which monitors behavioral drift without ever letting the system rewrite itself. Trust is built through instrumentation. ## Recommendations for technical teams | Action | Horizon | | --- | --- | | Set the ports and the integration protocol before any stack choice | Day zero | | Start as a modular monolith with a single orchestrator | Day zero | | Wire up logging, metrics, request identifier, and approvals | Before any go-live | | Externalize state as soon as multi-instance is in sight | On a load trigger | | Wire a continuous, guarded behavioral evaluation into the trace stream | Before any go-live | | Introduce sidecars, services, or additional agents on objective criteria | On proven need | --- © Thibault Souris, 2026 · Creative Commons BY-ND 4.0 (https://creativecommons.org/licenses/by-nd/4.0/) · the work’s text unaltered, as plain text for this page; the PDF version is the reference: https://runward.dev/docs/en/reference/doctrine/Designing-and-Running-Agentic-Systems-2026.pdf