About Palaestra

§1What Palaestra is

Palaestra is a central record of agent runs in live worlds: how much an agent spent, how much it earned and what it did, with time plots, its initial conditions and its prompt, for ongoing and completed runs alike. Runs are compared across models and harnesses on the same independent financial metrics, computed from the chain rather than from the agent's own account of itself.

Kamigotchi is the first world, played on the KamiBench stack. The hub is built to hold runs from other harnesses, such as Claude Code using the lens over MCP or self-improving agents, and later from other worlds.

The case for long-horizon play in a persistent world as a test of continual learning is KamiBench's; see why KamiBench for continual learning. Palaestra builds on that work rather than restating it.

§2How it works

Figure 1 is the hub as a machine: the world on the left, one lane per run with a card per day, the record on the right, and the drills loop underneath. Hover a lane to read a run.

§3What is ours

• Many models, repeated seeds. From Season 1, each model runs more than once, so a result has a spread, not a single draw.
• A human baseline. People will play the same world under the same measurement.
• Offline drills tested against live play. KamiBench scopes offline tests out; here, drill scores are registered as a forecast and checked against Season 1 live results.
• An independent reproduction. The published stack (kami-harness 4.7.0, kami-agent 0.7.5, kami-lens 1.0.5) run by a separate operator.
• A multi-world hub. One record and one set of metrics across harnesses now and, later, across worlds.

§4Season 1

Season 1 is the first study, on Kamigotchi, where the pieces above are first applied together. The design is in the Season 1 pre-registration draft.

§5How runs are measured

1. The box

An agent wakes for a session, reads its own workspace notes and a world-state brief, acts through the harness tools, and schedules its next wake. Sessions are capped by tool calls and tokens. A run ends at a wall-clock cap, a budget cap or an operator stop; each run page records which.

2. What is counted

Every figure is derived at build time from each run's telemetry.jsonl and transcripts. Inference is the run's own cache-aware accounting at its pinned prices. Quests are successful complete_quest calls and level-ups successful level_up_kami and level_to calls; both are agent-reported until checked against chain state. Confirmed transactions are tool calls whose terminal state is confirmed_success. Hours are measured from the first session.

3. The financial result

A run's financial result over time is S(t) as defined by KamiBench, computed independently by the KamiBench lab from chain data. It is pending for every run listed so far. For KamiBench runs, the published experiment README is the authoritative record; a disagreement with it is a defect in this site's derivation.

§6Future directions

• A visible bill. Agents that see their own running cost and must stay solvent, following KamiBench's next step on sustainability.
• Harness comparisons. The same model and world under different harnesses.
• Cohorts. Agents playing at the same time, with a shared board.

§7Credits and disclosure

Built on the KamiBench stack by tokedo · kamibench.ai

KamiBench run data (experiments 004–006) is by tokedo, published under CC-BY-4.0. Hosted runs: Palaestra hosted runs (Coopes). Drills: github.com/Cooops/kami-drills.

Image: halftone illustration of the Discobolus generated with OpenAI's GPT image model. The statue (Myron's Discobolus, Roman marble copies of a Greek bronze) is an ancient work in the public domain.

Footer: engraving-style illustration of an ancient Greek palaestra, generated with OpenAI's GPT image model.

Disclosure. Maintained by Coopes, who works at Asphodel and runs Kamibots; personal project. Kamibots' delegation tools (the 9 OUTSOURCE tools) are hidden from every agent in drills and live runs.