Strateji lobu, İcra lobu, Nöral Nöbetçi. The deep technical breakdown of how OSP composes production AI workers — with SİNAN as the canonical reference implementation.
Core Thesis: An AI Worker Is An Architecture, Not A Call
The most expensive misconception in enterprise AI right now is the idea that 'AI worker' means 'LLM call wrapped in a UI'. That framing is what produces the agents that fail in production — they have no concept of strategy separate from execution, no place for safety oversight that does not pollute the main reasoning prompt, no graceful degradation when a tool times out. Brain Architecture starts from a different first principle. An AI worker is a composition of three specialized loops, each with its own context discipline, talking to each other through a small structured contract. The architecture is the product. The model is just one component of it, replaceable on demand without redesigning the system around it.
Strateji Lobu, In Detail
The strategy lobe owns three responsibilities. Regime detection — figuring out what kind of situation the system is in. Is the user asking a quick factual question, a multi-step research task, an action request, an ambiguous message that needs clarification? Different regimes route to different downstream behaviors. Planning loop — given a regime, decompose the task into a sequence of tool calls and intermediate reasoning steps, with explicit fallbacks for likely failure modes. Fallback design — what happens when the catalog API times out, when the RAG returns nothing relevant, when the user changes their mind mid-conversation. The fallback is part of the plan, not an afterthought. Implementation-wise, the strategy lobe usually runs on a stronger reasoning model with low temperature, generates structured JSON plans, and is the easiest lobe to evaluate independently because plans are concrete artifacts.
İcra Lobu, In Detail
The execution lobe is the part that touches the world. Three subsystems. Tool registry — a typed catalog of every external action the system can perform: API calls, database queries, file operations, model-internal sub-calls. Each tool ships with a contract: input schema, output schema, expected latency, failure modes. The execution lobe never invokes anything outside the registry, which is what makes the registry the single audit point. RAG strategy — retrieval is itself a tool, but a special one. Two-stage retrieval is the default — semantic search to recall, reranker to precision. Hybrid retrieval, with keyword search alongside vector search, handles named entities better than pure vector. Action grammar — the rules for how multi-tool sequences compose. Idempotency where possible, rollback where not, structured error propagation back to the strategy lobe when a step fails.
Nöral Nöbetçi, In Detail
The neural sentinel is the lobe most teams skip and most regret skipping. Three categories of vigilance. Anomaly classes — outputs that fall outside expected distributions, conversations that exceed token budgets by 5x, tool sequences that loop, prompt injection attempts hidden in retrieved documents. The sentinel watches for all of them, often using cheaper models specialized on classification rather than generation. Kill switches — explicit triggers that stop the agent and route the conversation to a human. The kill switch documentation is what regulators want to see; the kill switch logs are what auditors want to read. Escalation paths — not every anomaly is fatal. Some warrant a slowdown, some warrant a clarification request to the user, some warrant immediate human handoff. The escalation matrix is a small table that lives next to the system prompt and gets reviewed monthly.
How The Three Lobes Compose
The composition contract is small on purpose. Strategy emits a plan as structured JSON. Execution consumes the plan and emits a sequence of (tool, input, output, status) tuples. Sentinel observes the stream of tuples plus the user-facing outputs, emits verdicts of {pass, slow, escalate, kill}. State flows through a shared scratchpad that is bounded in size — typically 8K tokens, aggressively summarized when it overflows. The flow is asynchronous; strategy can re-plan while execution is still running, sentinel runs continuously rather than at checkpoints. The orchestration code is small — usually under 500 lines — but the contracts between lobes are where reliability comes from. Bigger orchestration code is a smell that one of the lobes is doing the wrong job.
Eval Strategy Across Lobes
End-to-end evals are necessary and insufficient. They tell you whether the system is good; they do not tell you why when it gets worse. Per-lobe evals fill the gap. Strategy lobe — golden cases of (request, expected plan), scored on plan quality independent of execution outcome. Execution lobe — golden cases of (plan, expected tool sequence), scored on tool-use correctness assuming the plan was right. Sentinel — adversarial test set of inputs that should trigger {slow, escalate, kill} respectively, scored on classification accuracy. The full eval suite runs weekly. End-to-end evals run nightly on a smaller scenario set. Per-lobe evals run on every meaningful change to that lobe's prompt or model. The discipline is not glamorous; it is what separates a system that drifts silently from one that stays calibrated.
SİNAN As Canonical Example
SİNAN — the Archidecors customer-facing AI worker — runs Brain Architecture as described above. Strategy lobe routes between four regimes: quick question, design request, quote request, escalation. Execution lobe wraps a product catalog RAG, an image generation tool, an order-drafting tool, and a CRM integration. Neural sentinel enforces price-floor guardrails, flags suspected fraud, monitors for runaway-loop conversations, and owns the kill-switch handoff to a human salesperson. Eighteen months in production. Hundreds of conversations per day. Zero governance incidents. Eval pass rate measured weekly, hovering around 94% with model-version-specific dips that have always recovered after a one-week tuning pass. The architecture is replicable; the operating discipline is what makes it production-grade. Brain Architecture without the discipline is just a prettier diagram. With the discipline, it is the difference between an agent that ships and an agent that operates.