Case study
Route Agent
Multi-step workflows with a visible plan and a stop condition.
Plan, tool call, observation, and finish
- 4
- Graph nodes
- 6
- Step budget
- 0
- Hidden tool calls
Agents that hide their work
A hidden loop is difficult to debug and impossible to trust with a real inbox or database. Route Agent is a placeholder for an agent product whose plan and tool calls stay on screen.
Engineering approach
The planner emits a structured plan. Each act node can call one typed tool. The finish node may only cite observations already in state.
- JSON schema for plans and tool arguments
- A hard step cap plus a token cap
- Approval gate before any mutating tool
Challenges
The model would occasionally plan a tool that did not exist. Rejecting unknown tool names and asking for a revised plan removed that loop.
Results
Sample tasks finished inside the step budget with a trace a reviewer could follow. Replace this with your production numbers.
Technical implementation
A LangGraph state machine with plan, act, observe, and finish nodes. Tools are typed functions for search, SQL, and a calculator. A step budget ends the run.
- LangGraph
- Python
- FastAPI
- PostgreSQL
- OpenAI
Key features
- Explicit plan before any tool call
- Per-step token and time budget
- Tool traces stored with the answer
- Human approval on write actions
Written by
Anuoluwa Olutayo
AI Engineer