Long-horizon terminal agents can finish a task in a single run yet become unreliable when the same task is executed repeatedly. Stochastic errors propagate through actions and context, creating a gap between what the model can do and what it consistently executes.
RiseHarness is a framework that iteratively improves agent harnesses while keeping model weights fixed. It concentrates on task instances that share the same configuration but exhibit both successful and failed runs, using the successful executions as references for failure analysis.
A runtime monitor collects the agent's trajectory together with internal harness traces, allowing a fine‑grained examination of how model actions interact with command execution, feedback, context management, and completion decisions.
These observations are compiled into a Harness Handbook that links the evidence to specific harness behaviors and source code, facilitating the identification of recurring issues in mechanisms shared across tasks.
RiseHarness then searches other existing harness frameworks for comparable implementations and adapts them as design references for the current model and execution protocol. Candidate modifications undergo behavioral checks and repeated task evaluations; only versions that demonstrate measurable gains are fed into the next iteration.
This loop integrates suitable mechanisms from different frameworks into a single harness and continuously uses new execution evidence to steer further changes.
Evaluations across several models measure task success and reliability, ablate the contributions of harness execution evidence and external research, and test the transfer of improvements to unseen benchmarks with the same model.
RiseHarness treats a specific model as the unit of harness adaptation, studying how adjustments to shared interaction mechanisms can make existing capabilities more reliable across tasks.
Review