Agent architecture
Goal decomposition, tool schemas, working memory, context management and recovery from partial failure — designed as software, with typed interfaces.
Division 02
A chatbot answers. An agent completes. YTK AI builds systems that decompose a goal, call the tools they need, hold state across steps and verify their own output — then we instrument them so you can see exactly what they did, what it cost and where they will fail.
The reason most agent pilots never reach production is not model quality. It is that nobody defined what "correct" means precisely enough to test, nobody bounded what the agent is allowed to touch, and nobody costed a run at volume.
We treat those three as the actual engineering. Everything else follows.
Fig. 01 — orchestration graph with a verified retry path
The first four are what the system does. The last two are why it is allowed near a production process at all.
Goal decomposition, tool schemas, working memory, context management and recovery from partial failure — designed as software, with typed interfaces.
Specialised agents with bounded authority, an explicit handoff protocol and a supervisor that owns the final answer.
Ingestion, chunking, embedding strategy, hybrid search, permissions-aware retrieval and citation that points to a real source.
Purpose-built agents for a specific workflow — document processing, incident triage, research synthesis, back-office operations.
Golden sets, regression suites, adversarial cases, permission boundaries and full execution traces. If it cannot be measured, it does not go live.
Deployment, versioning, cost per run, latency budgets, drift detection and a rollback path that a non-researcher can execute.
Four stages, one loop. The verification stage is not optional — it is the difference between an agent that can be trusted with a workflow and one that quietly causes incidents at scale.
An agent that works 85 % of the time is not a success with a small caveat — it is a system with a defined incident rate. We quantify that rate before launch, watch it after launch, and design the human fallback for the remainder.
That is the honest shortlist. Agents earn their place when a human is doing something structured at volume, with a review step that already exists.
Extraction, classification, cross-referencing and exception flagging across long, inconsistent documents — with every field traceable to its source page.
Incident intake, categorisation, first-pass diagnosis and routing, with escalation when confidence or policy requires a person.
Multi-source gathering, contradiction detection and structured briefings that cite what they used rather than asserting from memory.
Permission-aware retrieval over the systems you already run, returning an answer and the evidence, not a link list.
Where YTK AI and YTK Robotics meet: agents that plan inspection routes, react to anomalies and coordinate multiple machines in the field.
We keep the orchestration layer independent of any single model vendor, so cost, latency and capability can be re-balanced without a rewrite.
We will tell you whether it is an agent problem, or a process problem.