Sign InOpen Brain
AI EngineerVideoSource Linked

Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic

Anthropic’s harness fixes for Sonnet 4.5 became harmful under Opus 4.5. Treat model-specific scaffolding as replaceable, and keep sessions durable enough to recover context and tool failures.

AI Engineer · Aug 11, 2026
Open Source Open MarkdownOpen JSON
Source Summary

**Sonnet 4.5** sometimes ended work early near its context limit, so Anthropic added context resets. With **Opus 4.5**, that behavior disappeared; the resets instead added latency and could discard cache state incorrectly.

Practical Implication

Keep harness components independently replaceable and retest every workaround when changing models. Persist session events outside the active context window so agents can restore discarded context, resume runs, and recover from tool failures.

Agent-Ready Context
**Sonnet 4.5** sometimes ended work early near its context limit, so Anthropic added context resets. With **Opus 4.5**, that behavior disappeared; the resets instead added latency and could discard cache state incorrectly.

Keep harness components independently replaceable and retest every workaround when changing models. Persist session events outside the active context window so agents can restore discarded context, resume runs, and recover from tool failures.

Claude managed agents shift loops, memory, and observability toward Anthropic while leaving context management and domain tools customizable. That reduces maintenance, but also places more behavior inside a managed surface that will continue changing with model releases.
Connected Context · Feed7 Judgment

This makes model upgrades an explicit harness-maintenance event: a context workaround can become pure latency or corrupt recovery assumptions when model behavior changes. It strengthens the candidates’ durable-state and deliberate-compaction patterns, but narrows them by showing that resets are not universally beneficial. The stable investment is replaceable controls plus external session history, not any fixed context policy.

Chained Recursive Language Models for Multi-Iteration ReasoningChained RLM deliberately resets context while preserving selected state; this Signal shows why that reset policy must remain replaceable and revalidated as model behavior changes.Anthropic's CCA Exam as a Field-Guide for Agentic Engineering — Frank Coyle, UC BerkeleyThe exam-derived checklist recommends deliberate compaction and stop-reason inspection; this supplies production evidence that those controls can become counterproductive after a model upgrade.Scaling to Long Horizons — Ross Taylor & Chengxi Taylor, General ReasoningBoth treat long-horizon context as a state-management problem, while this Signal adds the implementation consequence that session events must survive outside the active window for recovery.Codex, Behind the Harness — Dominik Kundel, OpenAICodex’s server-side compaction is a comparable managed context mechanism; Anthropic’s experience warns that such mechanisms need model-specific retesting rather than permanent assumptions.
Context Map
agentcoding#harness-engineering#context-engineering#agent-reliability
Uncertainty
Claude managed agents shift loops, memory, and observability toward Anthropic while leaving context management and domain tools customizable. That reduces maintenance, but also places more behavior inside a managed surface that will continue changing with model releases.