Claude Code 1M-token context + growth & token crisis
Key Questions
What is the context window size for Claude Opus and Sonnet 4.6?
Claude Opus and Sonnet 4.6 have a 1M-token context window available generally, with bi-weekly updates. This supports handling large inputs amid high usage of 350k daily active users and 1M pull requests.
What is context poisoning and how can it be fixed?
Context poisoning occurs when irrelevant or corrupted data degrades AI performance, making it 'dumber' over time. Fixes include proper context engineering, such as filtering inputs, using techniques like Chroma/RoPE/BEAM, temperature sculpting, and source-grafting as outlined in recent 101 guides.
What causes context decay in AI agents?
Context decay is a fatal flaw where AI loses coherence in long interactions due to dilution of key information. It stems from expanding context windows overwhelming the model's attention, addressed by structured workflows like sequential processing or split-merge patterns.
Why has Anthropic restricted OpenClaw access for Claude subscribers?
Anthropic blocked Claude Pro and Max subscribers from using flat-rate plans with third-party AI agent frameworks like OpenClaw due to a cost crackdown, as it led to 50x higher costs. This pushes users toward CLI and hybrid solutions with quotas and cache fixes.
What are the main workflow patterns for Claude Code?
Key patterns include sequential, operator, and split-merge workflows for efficient task handling. These evolved from prompts to harnesses, helping manage long contexts without economic nightmares like shoving 1,500 pages into windows.
What is the three-agent harness in Anthropic's design?
Anthropic's three-agent harness supports long-running full-stack AI development by coordinating multiple agents. It enables complex, autonomous applications while mitigating issues like context rot and security leaks.
How does rate limiting affect LLM applications like Claude?
Rate limiting manages unpredictable LLM traffic from high DAU, preventing overloads and ensuring stability. It's crucial amid growth, with quotas pushing users to CLI hybrids post-OpenClaw restrictions.
Why is the context window considered a crutch for AI?
Large context windows like 1M tokens act as a crutch by relying on brute-force inclusion rather than efficient memory. Alternatives include context engineering and agent harnesses to drop dependency and reduce costs.
Opus/Sonnet 4.6 GA (1M/bi-weekly) amid 350k DAU/1M PRs; new 101s on rot/decay/poisoning (Chroma/RoPE/BEAM, temp sculpting/source-grafting); quotas/cache fixes push CLI/hybrids post-OpenClaw (50x costs); workflow patterns (sequential/operator/split-merge); 3-agent harness; security leaks.