agon

or agony if you dont have enough VRAM.

Pit two AI agents against each other over any proposition and get a reasoned verdict. The proposition can be a design/architecture choice, a technical tradeoff, a strategy or policy, or a head-to-head X vs Y comparison.

The name comes from Ancient Greek drama

Motivation

I had an idea for using two agents to debate over a topic and produce a verdict. My first implementation was a python based MCP server you connected a harness to. I also tried this in go.

This worked, but I didn't like the idea of having to start the server each time, and connecting the harness afterwards before I could use this feature.

I attended GopherconUK 2026, and got particularly excited about the Kronk package: a go API that provided local inference with LLMs, with the intent to remove the need for a model server (or mcp server).

I wanted to experiment with using a spec document to guide an LLM on autopilot, rather than traditional prompting in a harness, to move from a MCP server architecture to a TUI app. I had a spec document and had Kimi K3 refine it. The spec details the requirements, constraints, and UX design choices I wanted. I sent Sonnet 5 on the implementation.

The result worked... but was quite bloated. I was happy with the general direction but saw a lot of room for improvement in terms of design, maintainability, and code quality (you can see so in the commit history).

Quick start

NOTE: This project is still in early stages of development.

go run ./agon

This opens the TUI at the main menu.

• Start a new debate: fill in the topic and (optionally) freeform starting context. You can specify paths to documents, code, etc, which will be added into the models read only sandbox.
• Browse archive: reopen any past debate, read-only.

Archived debates are written to debates/<session_id>.json (one file per session, written exactly once when the verdict completes; a killed or crashed run leaves no file).

Screens & keys

• Menu: ↑/↓ select · enter confirm · ctrl+c quit.
• New-debate form: tab/shift+tab move between fields · enter next field (or submit from the last field) · ctrl+s submit · esc cancel.
• Session view (live or archived; exclusive full-screen, no menu chrome while a debate is open): j/k or ↑/↓ scroll the transcript · a abort the live debate (with confirmation; discards the in-memory transcript, nothing is archived) · esc back to the menu (a live debate keeps running in the background; reopen it from the archive list, marked ●, to resume watching) · ctrl+c quit the app.
• Archive list: ↑/↓ select · enter open (read-only) · esc back to the menu.

Only one debate can be live at a time; starting a new one while another is running is blocked until you finish or abort it.

Debate modes

• proposition (default): one claim. advocate argues FOR, critic argues AGAINST.
• versus: two named options compete; each side advocates its own option.

Tones

formal (default), informal, genz, unhinged. Tone is style only; it never lowers the bar on substance, evidence, or honesty, and both sides share one tone so the debate stays even.

Evidence & debates

Debaters back claims with citations rendered as clickable markdown links in the transcript: [label](https://…) for external facts, [path:line](path#Lline) for files. If the starting context contains a path that resolves to an existing directory or file, both sides get sandboxed, read-only tool access (read_file, grep, list_dir) to that location so they can ground and cite real content; tool calls appear inline as collapsible ⚙ blocks and are recorded in the archive alongside the messages.

Makefile Targets

This project uses a Makefile with the following targets:

• make build - Compile the binary
• make run - Run the TUI in the foreground (Ctrl-C to quit)
• make test - Run all tests with the race detector, serialized (recommended for -race)
• make vet - Run go vet
• make fmt - Check gofmt formatting
• make clean - Remove the built binary