AgentShore: a coding agent alone vs. an orchestrated RL loop Left: a coding agent alone, where the human is the loop. Right: AgentShore's observe, decide, act, learn cycle running unattended, with the human setting direction and reading results. AgentShore An RL policy decides what to do next. CLI coding agents do it. TODAY A coding agent, alone You are the loop. …until you stop You write the prompt, pick the task Coding agent writes the code, opens the PR One patch one task, then idle Excellent inside a task. Nothing decides which task is next. vs AGENTSHORE An orchestrated loop You · goals, issues, guardrails RUNS UNATTENDED · OVERNIGHT 1 Observe Read the whole project state 252 signals — goal alignment, budget, agent health, failure history 2 Decide PPO policy picks one play 19 plays — pick up an issue, review a PR, run QA, groom the backlog, merge, prune 3 Act Hand it to a CLI coding agent Claude Code·Codex·Grok·Antigravity reviewer ≠ author, enforced 4 Learn Reward what actually closed project closure delta, not tokens spent — the policy updates before the next tick You · review the PRs, watch the dashboard, steer A coding agent decides how to write the code. AgentShore decides what to do next — and keeps deciding.