AgentRun Cuts Agent Costs by Learning Repeated Tasks

Agents waste tokens by repeatedly solving the same problems. AgentRun, an open-source platform, equips agents with a learning layer that captures and reuses task solutions. After each run, agents propose what to keep—notes, skills, or workflows—validated and stored in git. Subsequent runs inherit this knowledge, eliminating redundant work. The creators faced this cost spiral firsthand: code review agents exploded from $180 to $6,451 monthly. A custom agent trained on their own PRs costs 18 cents per review and catches 1.6x more bugs than the commercial service it replaced.

The platform wraps any major agent harness with durable execution, centralized monitoring, approval workflows, and spend controls. Server crashes don’t lose work; all runs persist in a central store. Agents can use Docker sandboxes, browsers, and Linux desktops. AgentRun supports both declarative workflows—typed steps, branching, tool calls—and full agent reasoning for hard cases. Learned knowledge lives in git, auditable and reversible.