Harnessing Claude’s Intelligence: Patterns for Evolving Agents

One of Anthropic‘s co-founders, Chris Olah, observes that models like Claude are grown more than built—researchers set conditions, but exact capabilities emerge unpredictably. This creates a persistent tension for builders: agent harnesses encode assumptions about what Claude cannot do on its own, yet those assumptions rapidly grow stale as the model improves. Even lessons from a few months ago need frequent revisiting. The article exposes the risk of accumulating ‘dead weight’ in agent code, where workarounds for limitations that no longer exist actually bottleneck performance.

The article lays out three concrete patterns. First, use what Claude already knows: starting with bash and text editor tools, which Claude composes into sophisticated behaviors. Claude 3.5 Sonnet reached 49% on SWE-bench Verified using only those two tools. Second, ask what you can stop doing—relinquish orchestration decisions to Claude. A code execution tool lets Claude filter, pipe, and transform tool results without streaming every byte through its context window. This lifted Opus 4.6 on BrowseComp from 45.3% to 61.6%. Similarly, skills enable progressive context disclosure, context editing removes stale content, and subagents let Claude fork into fresh windows. The article documents measurable gains across model versions: Opus 4.5 reached 68% on BrowseComp with compaction, while Opus 4.6 hit 84%. Third, set boundaries carefully via cache optimization (static-first prompt ordering, breakpoint management) and declarative tools for security/UX boundaries. Auto-mode in Claude Code uses a second Claude to judge bash commands, reducing the need for dedicated tools.

The core takeaway is that builders must continually re-evaluate their harness assumptions. The article’s vivid example: a long-horizon agent built for Sonnet 4.5 required context resets because the model had ‘context anxiety’ and wrapped up prematurely. With Opus 4.5, that behavior vanished—making the resets dead weight that would now harm performance. Serious builders should treat agent harnesses as living code that shrinks and simplifies as the model grows smarter. Pruning old constraints is not laziness; it is how you keep pace with a system that outgrows its own scaffolding.