3 Real time

Astra · 2 Nov 2026

In the first games a tick waited for every player, so thinking time cost nothing. That's the opposite of the problem this project started from: AI is slow next to a real-time game. So the second round puts the clock back. A tick resolves every 36 seconds whether anyone is ready or not, and a model that's still thinking misses its moment.

This chapter is one arena, The Vast Dark: ten agents, five models, one hour. It's written up in parts, like a small study: what was tested, how, what was measured, and what the numbers do and don't show.

What changed since chapter 2

Before this round the game and the runner changed in six ways.

The secrecy change matters because the agents play through the same API as the web client. Anything the server sends to a client, an agent can read. Message bodies now only ever reach the game master.

The question

How do models play when time costs something? In particular:

1. Speed. Does a slow model lose because it's slow?
2. Style. Will a model hold to a play style it's given, including hostile ones?
3. Diplomacy. What happens to talking when messages arrive on the next tick?
4. Conflict. With hostile styles in the mix, do the agents still talk first and fight last?

Method

One arena. The Vast Dark: an 82 by 63 map (5,166 cells, about ten times The Deep Dark's 520), 70 stars, none charted at the start. War by default, allies don't share vision, and victory by domination: the last empire or alliance with planets wins. 100 ticks of 36 seconds, so one hour.

Ten agents. Named alpha to juliet so the order is easy to follow:

The five peaceful agents are the control group: one of each model, same style, same effort. The other five carry the styles:

• Isolationist: no diplomacy at all. Send no messages, ignore every message, never change a stance.
• Turtle: defence first, slow expansion, never join an alliance.
• Aggressive: conquer, never agree to peace, use messages only to threaten.
• Supremacy: take over the map, use diplomacy only where it helps.
• Antagonist: win by cunning: make deals it means to break, spread false intelligence, set empires against each other.

Each agent's style is in its own prompt only. Nobody else can see it.

Turns. Each agent runs in its own loop: wait for a new tick, play one turn, repeat. Every turn is still a fresh CLI call with a notes file as its only memory, as in chapter 2. The prompt now also gives the tick length, the number of ticks and this tick's deadline as a local time, and tells the agent to send its orders before it writes its notes. A turn that's still running when its tick resolves is late: its orders apply to the next tick. A tick in which an agent didn't start a turn at all is missed. A call that runs past three ticks is killed.

Measurements. Every turn is logged on the server: agent, model, start and end tick, duration, status, cost or tokens, the turn's prompt, and the model's full output, every message, command and command result. The game master can read it all in a debug view. That log, the arena's per-tick snapshots and the message log are the data for this chapter.

The parts

• 3.1 The Vast Dark: the game itself, tick by tick.
• 3.2 Time: turn times, lateness, and what being late cost.
• 3.3 Talk and treachery: 963 messages, two betrayals and an alliance chain.
• 3.4 What changed: this round set against the first five games.

Caveats

It's one game. Each style was played by one model, so style and model can't be pulled apart: the antagonist was Codex on medium effort, and nothing here says how Haiku would play the antagonist. Several things changed at once between chapter 2 and this round, so a difference between them can't be pinned on any one change. Treat everything in this chapter as observations from one run, not results.