Breaking

Anthropic Publishes Harness Example for Long-Running Agent Loops

June 24, 2026 at 11:01 EDT

AI Agents
Software Dev & Coding
Foundation Models

In an engineering blog published on November 26, 2025, Anthropic detailed a harness implementation for "long-running agents" built on the Claude Agent SDK, arguing that iterative agent loops can steadily drive the long, messy tasks that single prompts struggle to handle.

December 2024 · Anthropic Engineering

The Rise of "Loop Engineering"

Developers are shifting from writing single prompts to designing automated agent loops — LLMs that repeat tool calls and observations to push through long, messy tasks. Costs rise sharply, but the results get measurably better.

2×

Longest session length doubled in 3 months

~45s

Median turn time, held steady

$100s

Monthly token spend under heavy use

Longest Claude Code Session

Sustained autonomous run time, measured over three months

<25 min

Start

→

>45 min

3 months later

The Agent Loop

Tool Call

→

Observe Result

→

Decide Next Step

↺

The model dynamically controls the loop — with triggers, stop conditions and validation harnesses set by the developer.

The Payoff

Better results on long, iteration-heavy tasks
Handles unpredictable coding, research, planning
A substantial share of internal code written by loops

The Cost

Token consumption swells substantially
Compounding errors & runaway loops
Quality drops when guardrails are weak

"Running a loop is easy. The hard part is the guardrails and validation."

Unlocking performance hinges on budget caps, context compaction and validation harnesses backed by evaluator agents — while many developers still get by with simple single calls.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.