TrajectoryRL
AI Agents That Cost Less and Work Better
An open competition where hundreds of participants optimize OpenClaw agent instructions.
Overview
How It Works
Hundreds Compete in Parallel
Participants write agent instructions for OpenClaw — the same runtime behind Claude Code. Hundreds explore the design space simultaneously, finding optimizations no single team could.
Tested on Real-World Tasks
Every submission runs through ClawBench: email triage, client escalation, standup prep — the kind of work AI agents do every day. Same tasks, same conditions, fair comparison.
Safety First, Always
Every safety and correctness check must pass. One failure disqualifies. You get cheaper agents without sacrificing the things that matter — no leaked data, no unauthorized actions.
The Best Pack Wins
Winner takes all. The cheapest agent that passes every check earns the reward. Today's champion gets challenged tomorrow — the bar keeps rising.
Scenarios
Where the Savings Come From
Smarter Instructions
Tighter prompts mean fewer tokens. Same results, 50-70% less cost — like editing a draft until every word earns its keep.
Fewer Tool Calls
Agents that check Slack once instead of five times. Clear constraints eliminate wasted round-trips.
Shorter Workflows
Why take 8 steps when 3 will do? Optimized workflows skip redundant processing and get to the answer faster.
Right Model, Right Task
Use Claude for judgment calls, a smaller model for lookups. Smart routing cuts costs by up to 93%.
Policy Packs
Plain English, Not Code
A policy pack is just a markdown file that tells an OpenClaw agent how to behave — the same format used by Claude Code. No training, no fine-tuning, no GPU.
| 1 | # AGENTS.md |
| 2 | |
| 3 | ## Identity |
| 4 | You are a senior executive assistant. |
| 5 | |
| 6 | ## Rules |
| 7 | 1. Safety > correctness > speed |
| 8 | 2. Never share confidential data |
| 9 | 3. Compress briefs to 90 seconds |
| 10 | |
| 11 | ## Tools |
| 12 | - slack: read-only unless approved |
| 13 | - exec: sandbox only, no network |
Stack
Architecture
Participate
Get Involved
For Miners
Write agent instructions that pass all safety checks at the lowest cost. No GPU, no training — just better prompts. The winner earns all the rewards.
Miner Guide →For Validators
Run the same OpenClaw + ClawBench evaluation pipeline independently. Your stake earns rewards for keeping the competition honest.
Validator Guide →