Blog

Updates, ideas, and release notes from the pi-go project.

Subscribe via RSS

Five coding agents, one protocol

Wiring Google Antigravity into the Agent Client Protocol sounded like a one-afternoon job. It wasnt — and the reason is what supports ACP actually means.

Read post →

Memory Palace: four layers of context

pi-go Memory Palace combines SQLite facts, semantic embeddings, a temporal knowledge graph, and background project miners that learn while you code.

Read post →

Running Mistral in a multi-provider CLI

Adding the Mistral provider to pi-go was a series of small, sharp lessons — each one a bug that only shows up against the live API.

Read post →

Subagents: fanning out work without losing the thread

pi-go subagent system dispatches typed roles — explore, plan, designer, reviewer, task — over the Agent Client Protocol, with worktree isolation and concurrency guardrails.

Read post →

The DeepSeek and GLM-5.3 effect

Switching pi-go default model to DeepSeek and adding GLM-5.3 tripled my agent usage. A data-driven look at the session history.

Read post →

Your coding agent should draw diagrams

I taught pi-go to draw Mermaid diagrams in the terminal. The interesting part turned out to be everything that isnt the diagram.

Read post →

OpenRouter arrives in Pi-Go

Pi-Go now supports OpenRouter as a first-class provider. Select models with the openrouter/ prefix and use tool calling, streamed responses and reasoning tokens.

Read post →

Introducing Pi-Go

pi-go is a terminal-based AI coding agent built on Google ADK Go with multi-provider LLM support, sandboxed tools, and Memory Palace.

Read post →