Storey 05 / 05 / The Pagoda
Agents
An agent is a language model that can act. It calls tools, reads files and runs commands, as well as answering.
I build agent workflows with LangGraph and connect them to tools over MCP. When data must stay private, I run the model locally with Ollama. I build software with Claude Code, often as a team of agents. I score what models write with FActScore and G-Eval.
The main thing I have learned is to keep the model away from the numbers. Code works out the figures and the model explains them. Then a check makes sure every figure it states came from the data.
The tower
Tools 7
- LangGraphThe finance app's Assistant is one LangGraph graph that sends each question to a small set of read-only tools. Also used for agent workflows at PA Consulting.Finance Agent →
- MCPA server that lets AI assistants ask the deforestation project for its results.Deforestation Detection →
- Claude CodeEvery agent in the software factory is a Claude Code session. On the internship it wrote the tests first, under strict test-driven development.Software Factory →
- OllamaRuns the finance app's model on the Mac itself, so questions about money never leave it.Finance Agent →
- FActScoreChecks each fact in the football commentary against the match data.Football Intelligence →
- G-EvalHas a model grade the football commentary against set criteria.Football Intelligence →
- sub-agent workflowsThe deforestation project was built with Claude sub-agents.Deforestation Detection →
Notes from this storey
Notes the model never saw
My finance Assistant's local model sent £50.00 as 50 pence and guessed at category names. The notes on each tool's inputs were code comments, which a model never sees.
Memory that kept growing
On a 16 GB Mac, the default Ollama builds of two models used more memory the longer a chat went on: about 12 GB after 25 questions.
Standing on this storey
Project / 2026
Software Factory
A team of AI coding agents that takes a software job, plans it into pieces and builds the pieces at the same time.
The team needed between 39% and 64% of the time one agent took alone.
Read the chapter