Agentic Engineering & ToolingTools-in-Action30min
One tool to rule them all: bash, a tiny LLM, and the birth of a coding agent
A live demo shows how a tiny local LLM becomes a coding agent using just a bash tool and a simple loop. It explores tool calls, prompt engineering, error recovery, and why one built-in tool can outperform many servers—revealing how coding agents work under the hood.
talk.summaryAiDisclaimer
Philippe CharrièreDocker
talkDetail.whenAndWhere
Monday, October 5, 18:20-18:50
TBA 5
talks.roomOccupancytalks.noOccupancyInfo
Coding agents look like magic. Strip away the hype and you find a surprisingly small core: a loop, a language model, and… a shell.
In this Tools in Action, we will see this core live. We give a small, local LLM (under 10B, running on Docker Model Runner) a single built-in tool (bash) and watch it turn into a coding helper. It lists files, greps the repo, runs the tests, curls an API, pipes into jq… all on its own, because the model already knows these CLIs. And when it doesn't? It runs --help and figures it.
We'll see the smallest agent loop in action: the model decides which command to run, the agents executes it, feeds the result back, and repeat until it answers. We'll look at what actually goes over the wire (the tool description, the tool_calls the model emits), why writing a good description is really prompt engineering, and why returning errors is what lets the model recover on its own.
You'll leave understanding how coding agents really work under the hood, why one well-chosen built-in tool often beats a pile of MCP servers, and how far a tiny local model can go.
Live demos, few slides, code you can run tonight.
In this Tools in Action, we will see this core live. We give a small, local LLM (under 10B, running on Docker Model Runner) a single built-in tool (bash) and watch it turn into a coding helper. It lists files, greps the repo, runs the tests, curls an API, pipes into jq… all on its own, because the model already knows these CLIs. And when it doesn't? It runs --help and figures it.
We'll see the smallest agent loop in action: the model decides which command to run, the agents executes it, feeds the result back, and repeat until it answers. We'll look at what actually goes over the wire (the tool description, the tool_calls the model emits), why writing a good description is really prompt engineering, and why returning errors is what lets the model recover on its own.
You'll leave understanding how coding agents really work under the hood, why one well-chosen built-in tool often beats a pile of MCP servers, and how far a tiny local model can go.
Live demos, few slides, code you can run tonight.