Anthropic publishes long-running agent harness playbook

// 71d agoTUTORIAL

Anthropic publishes long-running agent harness playbook

Anthropic’s November 26, 2025 engineering post explains why long-running coding agents fail across context windows and proposes a two-agent harness (initializer plus coding agent) that preserves continuity with artifacts like `init.sh`, feature lists, progress logs, and git commits. The guidance is grounded in Claude Agent SDK and Claude Code workflows, arguing that execution structure matters more than adding tools blindly.

// ANALYSIS

Harness engineering is becoming the real moat: better session design can outperform model-only upgrades on long, branching tasks.

–Breaking work into small feature increments helps avoid one-shot builds that collapse mid-context.
–Persistent handoff artifacts (progress file, feature JSON, git history) let new sessions recover state fast.
–Session discipline (run `init.sh`, verify core flows, then implement) catches regressions before new code piles on.
–External community projects and plugins are already operationalizing Anthropic’s pattern, which suggests broad practical demand.

// TAGS

claude-codeagentai-codingprompt-engineeringtestingdevtool

DISCOVERED

71d ago

2026-03-17

PUBLISHED

71d ago

2026-03-17

RELEVANCE

8/ 10

AUTHOR

Prompt Engineering

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE2h ago

Cursor adds dedicated subagents for skills

Cursor now allows developers to execute tool-heavy or research-intensive agent skills within dedicated subagents. This architectural shift isolates noisy background tasks, keeping the main chat context clean and focused.

UPDATE2h ago

YouTube moves AI labels to video player

YouTube is moving its AI content disclosures from video descriptions to more prominent placements beneath the player and on Shorts overlays. Starting in May, the platform will use internal signals to automatically label photorealistic AI content that creators fail to disclose.

OPEN SOURCE6h ago

Taste Skill kills AI "frontend slop"

Taste-Skill is an open-source framework that provides portable "agent skills" to enforce high-end design principles in AI-generated code. By injecting specific design directives and "anti-slop" rules, it enables LLMs to produce editorial-grade UIs that bypass generic, boilerplate-heavy AI templates.