# Long-running Agents

> **TL;DR.** Long-running agents are AI agents designed to sustain work across multiple context windows, persisting state through structured artifacts — progress files, git commits, feature specs — so each new session resumes where the last ended.

- **Category:** AI / Developer Tools / Infrastructure
- **Stage:** established
- **Age:** 202 days
- **Origin date:** 2025-11-26
- **First detected:** 2026-04-30
- **Canonical URL:** https://earlyterms.com/term/long-running-agents
- **Sources:** 7 primary URLs

## Definition

Long-running agents are AI agents designed to sustain work across multiple context windows, persisting state through structured artifacts — progress files, git commits, feature specs — so each new session resumes where the last ended. The pattern addresses a hard constraint: every context window is amnesia.

Anthropicエンgineer Justin Young formalized the concept on [November 26, 2025](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents) with a two-agent harness (initializer + coding agent) enabling Claude to build production web apps across sessions. By February 2026, Cursor launched a research preview with documented 36–52 hour autonomous coding runs; Anthropic followed with a [three-agent planner-generator-evaluator architecture](https://www.anthropic.com/engineering/harness-design-long-running-apps) in March 2026.

## Example

Cursor's long-running agent completed an all-new chat platform integration in 36 hours and a mobile app port in 30 hours, producing PRs with merge rates comparable to short-session agents. The runs used multi-agent plan-and-verify loops to prevent context drift on tasks spanning thousands of files.

## Analogy

Think of it as a relay race for AI: each runner picks up the baton from a structured handoff note, not from memory.

## Why it's emerging now

Frontier models can now sustain 25–52 hour autonomous coding sessions with proper harnesses. Anthropic's November 2025 engineering post established the canonical two-agent pattern; Cursor's February 2026 research preview demonstrated it at production scale (151k-line PRs). Three forces converge: models capable enough to stay coherent, harness patterns proven at scale, and token costs low enough to run for days.

## Related terms

- *parent:* agent-harness
- *related:* managed-agents
- *related:* context-engineering
- *related:* context-rot
- *related:* agent-loop
- *related:* parallel-agents
- *parent:* coding-agents
- *related:* agentic-coding
- *related:* cloud-coding-agents
- *related:* context-window
- *related:* harness engineering
- *alias:* multi-session agents

## Sources

1. [Anthropic Engineering — Effective Harnesses for Long-Running Agents (Nov 26, 2025)](https://www.anthropic.com/engineering/effective-harnesses-for-long-running-agents)
2. [Anthropic Engineering — Harness Design for Long-Running Application Development (Mar 24, 2026)](https://www.anthropic.com/engineering/harness-design-long-running-apps)
3. [Cursor — Expanding our Long-Running Agents Research Preview (Feb 12, 2026)](https://cursor.com/blog/long-running-agents)
4. [Cursor — Scaling Long-Running Autonomous Coding (Jan 14, 2026)](https://cursor.com/blog/scaling-agents)
5. [Hacker News — Effective harnesses for long-running agents (125 points)](https://news.ycombinator.com/item?id=46081704)
6. [Amplify Partners — How Hightouch Built Their Long-Running Agent Harness (Jan 20, 2026)](https://www.amplifypartners.com/blog-posts/how-hightouch-built-their-long-running-agent-harness)
7. [Addy Osmani — Long-running Agents (Apr 28, 2026)](https://addyosmani.com/blog/long-running-agents/)

---
_Generated by EarlyTerms · https://earlyterms.com/term/long-running-agents_