Long-Running Agent Systems
Long-running agents shift the central engineering problem from a single answer toward handoffs, task decomposition, evaluation, concurrency, environment reliability, and merge coordination. The corpus supports no claim that they are autonomous in the strong sense: all examples retain tests, evaluation, review, or explicit recovery mechanisms.
Synthesis
- Anthropic’s harness examples use planners, generators, evaluators, structured handoffs, browser testing, containers, task locks, and CI to extend work beyond one context window. [src] [src]
- Ralph-style loops and Cursor cloud agents show simpler operational forms: repeat bounded work, isolate environments, return artifacts, and make completion or review visible. [src] [src]
- Wheelhouse’s work-graph vision highlights a scaling tension: parallel workers can accelerate implementation, but shared state, merge queues, and capacity constraints become first-class coordination problems. [src]
Coverage expansion
GitHub's Copilot plan changes make capacity part of the design problem: the company attributes stricter session and weekly limits to compute demand from long-running, parallel agentic workflows, and recommends reducing concurrency or choosing lower-multiplier models near limits. This is vendor-specific operational guidance, not a general measurement of agent efficiency. [src]
The workspace and recovery side of parallel execution is developed in Version Control And Agent Workflows.