Multi-Agent on Corey Daley

Before the First Commit: What Multi-Agent Sprint Planning Actually Catches

Sat, 11 Apr 2026 18:05:00 -0400

What does a multi-agent sprint planning workflow actually produce? Not just a cleaner document — it finds real bugs in a plan before implementation begins. When the /sprint-plan command ran against a “simple” Go REST API project, the security review phase returned three critical findings: a logical contradiction that made the stated auth behavior impossible, a schema constraint that would silently break token revocation, and a SQLite pragma applied to only one connection in a pool.

The post walks through the entire planning session for org-api — from seed prompt to approved sprint document — showing what each phase of the review pipeline produced and what changed as a result. The security findings came from reading the plan carefully, not from running any code. That’s the point.

What step in your planning process is explicitly there to prove the plan wrong before implementation begins?

From Config Hub to Competing Voices: How agent-config Became My AI Collaboration Stack

Sat, 11 Apr 2026 14:50:00 -0400

I started agent-config as a shared configuration hub: one repository to rule Claude, Codex, Copilot, and Gemini. That lasted about two iterations before the cracks showed. Forcing every AI agent to share the same configuration format was the wrong abstraction — different tools, different philosophies, different file formats. The solution wasn’t more uniformity. It was a different model of collaboration entirely.

Today agent-config is Claude-specific, but Codex is still central to how I work. The difference: Codex is no longer a configuration target. It’s a competitive collaborator. Sprint plans, blog posts, security audits — every significant output runs through a workflow where Claude and Codex produce independent drafts, critique each other’s work, and force synthesis from the tension. Two AI voices with different instincts produce better output than either would alone — just like a team of people with different backgrounds does.

Is your multi-agent workflow built for sharing configuration, or for generating the productive disagreement that makes output actually better?

Polyphon at v0.8.0: The End of the Prototype Phase

Mon, 23 Mar 2026 19:55:00 -0400

When I shipped Polyphon v0.1.0-alpha.2, the pitch was simple: put multiple AI voices in one conversation and let them respond to each other. That was useful. But early usefulness and long-term trust are not the same thing.

v0.8.0 is the release where Polyphon crosses that line. The features that made the difference weren’t the ones I planned at launch. Voices can now interact with actual files, with per-voice sandboxing and explicit permission categories. Conversation history is encrypted with SQLCipher whole-database AES-256, with optional password protection. FTS5 search turns the archive into working memory you can actually retrieve from. These aren’t incremental improvements — they’re the features that decide whether a tool stays an interesting experiment or earns a place near real projects.

What actually makes you trust an AI tool with real work: capability, privacy, or memory?

Polyphon's MCP Server Makes Your AI Ensemble a Native Agent Tool

Mon, 23 Mar 2026 19:35:00 -0400

Polyphon just shipped MCP server support, and it reframes what the tool actually is. You’ve been opening it as a UI. Now your coding agent — Claude Code, Cursor, Codex CLI, GitHub Copilot — can call into it directly and broadcast a prompt across your entire Claude + GPT-4o + Gemini ensemble in a single tool call.

The practical win is not just convenience. It is that you can package disagreement: three independent models, each with different priors, applied to the same question at once. Five tools, two CLI flags, one Settings toggle. The SessionEventSink abstraction is why this works without duplicating orchestration logic — desktop UI and headless agent calls run on the same engine.

Where in your workflow would paying for disagreement actually change a decision, rather than just produce a louder answer?

I Built a Tool So AI Models Could Talk to Each Other

Mon, 16 Mar 2026 12:30:00 -0400

Every AI power user I know runs the same manual workaround: ask Claude, ask GPT, copy the interesting parts of each into the other, then try to synthesize what you learned. The models are good. The coordination is not.

I just shipped Polyphon v0.1.0-alpha.2 — a free, local-first desktop app that puts multiple AI voices in the same conversation so they can actually respond to each other. You’re the conductor. They’re the ensemble. Save a group of voices as a composition and reuse it whenever you need that ensemble again.

What should a multi-agent conversation feel like when you’re not building a pipeline — when you just want to think out loud with several models at once?