Six focused tools in one web app for work that used to be spread across several systems.

Before the Hub, the team moved between systems by hand. The app now provides one place to use several internal APIs and has substantially reduced that manual work. It covers duplicate handling, catalogue search without SQL, coordinate-based destination lookup, bulk processing and mismatch inspection.

Some tools use an LLM to evaluate candidates. The model returns a structured suggestion with a one-line reason that the reviewer can check. A person still makes the decision.

The bulk tool handles a full team workflow. Users upload a batch and assign parts of it to team members. Progress streams to the browser, and failed operations can be retried without running the whole batch again.

An agent platform for the same operations work, running on a tool-calling runtime I wrote myself.

The Hub still needs someone to open it and drive each tool. The Agent starts when work arrives and handles the job from there. Chat, Jira, Slack and scheduled tasks all feed into the same runtime. Sessions can use about 48 domain tools. At runtime they load named skills that provide instructions and grant access to the tools a job needs, so each session only gets the write access tied to its skills.

Every turn writes an immutable, cache-aware cost record in nano-USD, which gives the exact cost of each ticket. The project has 103 test files and 10 database tables.

Four tools give AI agents across the company access to around 300 GraphQL operations.

The team that owns the graph asked me to help because of my agent engineering and LLM experience. I designed the tool schema and helped set the direction. I did not write the server's production code.

Giving every operation its own tool would fill the context window before the agent started working. Instead, the agent follows four steps: list operations, ask what one does, inspect its schema, then run it. The graph can grow without adding more tools. Business context has its own step because an operation name alone rarely tells a model whether it is the right choice.

The server is in production, and employees use it from their own agent applications.