AAI News Hub
ResearchWed, August 5, 2026·Aug 52 sources corroborating

Papers target better training data for terminal agents

RST and CalibForge synthesize verified terminal tasks for long-horizon agent training.

Why it matters

The work addresses a key bottleneck in agent development: producing executable, verifiable tasks that are both consistent and useful for learning. It suggests terminal-agent training may rely more on automated task generation and solver-calibrated curricula than expensive manual authoring.

The key points

  • 1.RST generated 37,484 verified terminal tasks across 15 recursive rounds.
  • 2.CalibForge produced 5,431 tasks calibrated against solver behavior.
  • 3.Both papers focus on scalable supervision for long-horizon terminal agents.

Two new papers describe systems for generating training tasks for terminal agents. Recursive Synthetic Terminal Tasks (RST) recursively expands verified seed tasks, producing 37,484 synthesized tasks across 15 rounds at roughly $0.05 per task, with later rounds becoming harder by solution length, command count and solver pass rate. CalibForge builds 5,431 calibrated terminal tasks by using verified solver behavior to revise candidates toward learnable difficulty, with reported gains on Terminal-Bench 2.0, SWE-bench Pro and Doc2Repo.

Try this today

Read the papers before building terminal-agent datasets, and add verifier-backed validation plus solver calibration to your data pipeline.

Sources & original reporting

This brief summarizes and links to reporting from the publishers below.

Enjoyed this brief? Get the next one in your inbox.

More in Research