Anthropic released Claude Opus 4.7 with a million-token context window and the ability to run autonomous coding and research tasks for up to twelve hours. For the first time, a frontier model can hold an entire mid-sized codebase in its working memory and act on it without a human re-prompting every few minutes.
The practical implication is bigger than the benchmark gain. A year ago, agent products needed elaborate retrieval pipelines, summarisation chains, and prompt rewrites to survive past a few thousand tokens. Most of that scaffolding is now an anti-pattern. Pull the codebase in, point the agent at the goal, and let it work.
We migrated three internal automation tools off custom retrieval in the last six weeks. The simpler stacks ran 38% faster end-to-end and produced 12% fewer broken outputs. The lesson isn't that retrieval is dead — it's that we were retrieving around a constraint that no longer exists for code-bounded tasks.
The new constraint is observability. A twelve-hour run can quietly burn through a budget or wander into the wrong subgraph. Build tracing into the agent's tool calls from day one. Sample completed runs the way you'd sample production traffic. Treat each long run as a workload, not a prompt.
If you're starting a new agent product in mid-2026, you should not be writing retrieval logic before week three. Start with a long-context call, instrument heavily, and only add structure when the model demonstrably needs it.
