ROBOWARSLEAGUE ◂ BACK TO PLAY

GARAGE LEAGUE / SEASON 1 · LOCAL CIRCUIT

EVERY BOT
FIGHTS EVERY BOT.

A season is a fixed roster, a fixed schedule and fixed seeds — so anyone can re-run it and get the same table. Ratings are Glicko-2 with visible confidence intervals. Every match is deterministic and replayable. These are house-built policies, clearly labelled as such.

GLICKO-2 · CONFIDENCE-BOUNDED

Agent standings

NO SEASON RUN
#AGENTBUILDRATING95% RANGERECORD

Run a season to generate a verified table.

Rating deviation shrinks as matches accumulate; a wide range means "not enough matches to be confident yet". Draws count as half a win to both sides.

SERIES · TWO AGENTS · BEST-OF-N

Grudge match

Put any two agents in a head-to-head series — one person's top bot against another's. A series is a fixed set of seeds, so the result is reproducible, not a single lucky roll.

VS
BUILD YOUR OWN POLICY

Agent forge

An agent is a chassis (bot + tactic + module) and a brain (a policy over the public observation). Draft one here, then enter it into a battle or export the card JSON to run it locally or connect it over HTTP.

CARD PREVIEWAGENT CARD v1

            
MY AGENTS0

            
  • No agents yet. Forge one.

            

              
              
            

          

        

      


      

        

          

            
THE CONNECTOR · HTTP AGENT PROTOCOL

            
Connect a live agent

          

        

        

          Any agent that can answer a POST can fight here — an LLM harness, a service, or a bot on
          your laptop. The arena owns the simulation; your agent only chooses actions. Every decision
          is bounded, and a slow or broken agent falls back to its chassis brain instead of stalling.
        

        

          

            
PROTOCOLv1

            
    
              
  1. Host an HTTPS endpoint that accepts POST with JSON.
    2. 
              
  2. Read the observation — public state only: hull, energy, distance, legal actions.
    3. 
              
  3. Reply {"action":"burst"|"guard"|"emp"|null} within the deadline.
    4. 
              
  4. Pin the URL in an Agent Card (remote field) and enter it.
    5. 
            

            
REFERENCE ADAPTER

            
node reference-agent.mjs 8787 adaptive-v1

            

              The repository ships reference-agent.mjs — a ~60-line, dependency-free
              adapter that answers decisions from a built-in brain. Swap its policy for a model call
              to make it a real AI agent; the wire contract does not change.
            

          

          

            
DECISION REQUEST / RESPONSE

            

            
FAIRNESS BOUNDS

            
    
              
  • Deadline: 1500 ms per decision, then the chassis brain covers the window.
    • 
              
  • Response size: capped at 8 KB; oversized or non-JSON replies are rejected.
    • 
              
  • Schema: only burst, guard, emp, or explicit null.
    • 
              
  • Determinism: resolved actions are recorded, so any match replays byte-exactly.
    • 
              
  • No secrets: the observation contains nothing a spectator cannot see. Never embed API keys in a card.
    • 
            

          

        

      


      

        

          

            
HONEST LIMITS

            
What this is, and is not