agents
Automation driven by a model, and the guardrails that make it safe to run.
-
Tracking Meals and Automating Shopping With an MCP Server
An MCP server for meal tracking and grocery shopping automation: planned meals minus live stock, and why the v2 rebuild derives stock from an event journal.
-
Treating Issue Bodies as Untrusted Input
Prompt injection via GitHub issue bodies: fence untrusted text and accept model output only as a fixed JSON action schema, tested without a model.
-
From a Large Allowlist to Three Denials: Permissions for a Coding Agent
bwrap --unshare-net --cap-drop ALL instead of a command allowlist: sandboxing a coding agent's shell so rm -rf and curl cannot leave the workspace.
-
Self-Hosted LLM Inference: Serving, Benchmarking and Agent Guardrails
Ollama OpenAI-compatible /v1/chat/completions: a repeatable benchmark on your own tasks and a dry-run plus --approve guardrail for infrastructure agents.
-
Verifying an Agent's Work Against Reality, Not Its Own Report
git diff --quiet HEAD and a re-run test command in a separate CI job: verifying a coding agent's work from repository state instead of its own report.
-
Testing an Agent Harness Without Ever Calling the Model
Unit testing a coding agent's tool-call permission layer with pytest and recorded fixtures: allow, deny or ask decisions with no model or API key.
-
The Label as Contract: Consent and Priority as the Whole Queue
agent-ready label as the only consent gate for a coding agent: FIFO order by label time and a re-read-then-swap claim so two runs never share an issue.
-
A Deterministic Daemon That Turns a Labelled Issue Into a Pull Request
git apply --check before git commit: an issue-to-PR daemon where the model only returns a patch and plain Python commits, pushes and opens the PR.