Agentic Flow

The Knowledge Flow turns one source into a graph of facts, concepts, and syntheses. The Agentic Flow is what an agent does on top of that graph — it is not a fixed sequence. An agent researches, queries, and authors reports in whatever order the task demands, looping back to fetch more sources whenever the graph is thin.

Three phases

1. Research — gather sources around a topic and feed them into the repository. See Research.
2. Query — read what the system knows: search facts and concepts, read syntheses, walk related concepts. See Query.
3. Reports — author a report; OKT annotates every sentence with the facts it rests on. See Reports.

Each phase is a set of MCP tools — the agent's verbs. The order is up to the agent.

How the phases compose

The three phases are not a pipeline. An agent typically:

1. Starts with research — searches for sources, opens an investigation, fetches them.
2. Polls getSourceTasks until the pipeline drains.
3. Queries the graph — searches facts and concepts, reads syntheses, walks related concepts.
4. If a query surfaces a thin area (few facts, no synthesis), loops back to research with a sharper query.
5. Once the graph covers the topic, creates a report and reads it back annotated.

The loop — research → query → (thin?) → research → ... → report — is the core agentic pattern.

Building agents for OKT

OKT exposes its capabilities as MCP tools, so any agent framework that speaks MCP can drive the agentic flow. The phases map cleanly onto custom agents — one agent per phase, or one orchestrator that does all three. The phase pages describe the responsibilities of each phase so you can build agents against them in whatever framework you use.

A common pattern is to build three specialized agents — a researcher, a querier, and a reporter — each with a tight prompt scoped to its phase's tools, plus a primary agent that orchestrates them. Keeping each agent narrow (one phase, one toolset, one prompt) makes the loop easier to debug and the prompts easier to tune.