Five working agent scaffolds you can clone and run on your laptop with a local Ollama install. No OpenAI key. No Anthropic key. No vendor lock-in to demonstrate. The point is to see the composition : how agents share state, hand off to each other, validate inputs, and produce structured output : without the friction of paid API setup.
System Core Architecture
All 5 Scaffolds : Blackboard Pattern + LLM Router
The five scaffolds
Code reviewer. Takes a diff, runs a multi-pass review, flags concerns by severity. Useful as a pre-commit hook on a personal project.
Blogger. Outline → draft → editing pass → ready-to-publish markdown. Mostly an exercise in composing agents around a writing pipeline; quality of output is bounded by the local model you point it at.
Bug triager. Reads a freshly filed issue and a small slice of repo context, suggests labels, severity, and a likely owner. Lifted from real ServiceNow triage work I did in the L2 role.
Doc summariser. Take a long doc, produce structured summaries at three different levels of compression. Demonstrates the “structured output at every boundary” pattern.
Schema migrator. Given an old and new database schema, drafts a migration plan. The least mature of the five : included as the example of “here's where the pattern stops working without a stronger model.”
The Blackboard pattern in one sentence
Instead of agents calling each other directly, they all read from and write to a shared structured state object. Each agent declares what it reads, what it writes, and what preconditions must be true before it fires. The orchestrator picks the next agent to run based on the state.
The advantage is testability : each agent can be exercised in isolation with a stubbed blackboard. The disadvantage is that you have to be disciplined about the schema. The repo includes the schema definitions and a few worked examples.
No paid API integration. You can wire one in : the abstraction is there : but it's not the point.
No fine-tuned models. Everything runs on whatever you pull into Ollama.
How to evaluate the code
Clone the repo. Install Ollama. Run the code reviewer against a small diff in any of your own repos : it will produce structured output you can scroll through. If the composition is clear from reading two scaffolds end-to-end, the track has done its job.