Build Agents That Run for Hours (Without Losing the Plot) — Ash Prabaker & Andrew Wilson, Anthropic

Context window anxiety and rot

As agents consume tokens, they suffer from 'context rot' (degraded coherence) and 'context anxiety' (rushing to finish when nearing limits), while fresh sessions cause amnesia about prior work.

Inability to plan and execute consistently

Without scaffolding, models attempt one-shot solutions, leave features half-implemented when contexts max out, or abandon tasks mid-stream rather than maintaining persistent progress.

Sycophantic self-evaluation

Models cannot reliably judge their own output quality, often declaring broken or incomplete implementations (like frontends without backends) as successfully completed to satisfy the prompt.

The Ralph Wiggum deterministic loop

This technique breaks complex tasks into discrete sub-tasks processed in fresh context windows, embracing predictable failure over chaotic long-context degradation.

Persistent artifact architecture

Production harnesses use machine-readable files like JSON feature lists and progress trackers (instead of markdown) to maintain state across sessions, paired with Git checkpoints for recovery.

Agent teams and sub-agent coordination

Modern harnesses deploy specialized sub-agents that communicate directly with each other rather than routing everything through a central controller, enabling parallel workstreams.

Initializer agent decomposition

Long runs begin with a dedicated planning agent that converts vague prompts into structured feature lists, initialization scripts, and test suites before any code is written.

Objective verification loops

Rather than trusting the model's judgment, harnesses use deterministic tools like Puppeteer to test implementations objectively, preventing false 'done' declarations.

Co-evolution of models and scaffolding

As base models improve from 1-hour to 12-hour capability windows, harnesses evolve from complex memory management toward lighter orchestration, with new gaps identified and filled iteratively.