Better Models Shift Advantage to the Runtime

The strongest July 25 signal came from Digg AI and Digg Tech surfacing Anthropic’s Claude Opus 5 launch, then from Anthropic’s own release notes and operator guidance explaining why the gain matters. The important pattern is not just a better coding model. Anthropic says Opus 5 approaches frontier capability at half the price, converts extra effort into better results more reliably, verifies its own work without being told, and benefits from stripping away old prompt scaffolding. Thariq Shihipar’s same-day context-engineering guide makes the shift explicit: Claude Code removed over 80% of its system prompt for advanced models with no measurable loss on coding evals. Boris Cherny then highlighted a second-order implication surfaced in the system card and red teaming: stronger models are also becoming harder to prompt inject. Together, these signals point to a product change in where quality comes from. More of it now comes from runtime design than from hand-written prompt ceremony.

Many teams still treat agent quality as a prompting problem. That is getting less true. As models become more proactive, more capable at self-verification, and more expensive to over-instruct, the bottleneck moves outward into system design. Builders who keep layering on verbose prompts, duplicated verification text, and rigid subagent rules will pay more and often get worse behavior. Builders who treat context assembly, effort controls, model routing, tool permissions, and injection resistance as first-class product surfaces will ship faster loops with clearer failure modes and lower operating cost.

Refactor one production agent into a runtime contract instead of a giant prompt. Trigger: a user asks for a multi-step task such as debugging, account analysis, onboarding ops, or document generation. Context: load only the repo, customer record, docs, or artifacts needed for this run, plus stable memory and policy surfaces that do not belong in the live task prompt. Tools: expose the minimum tool set required, define when subagents are allowed, and route flagged or lower-stakes work to cheaper fallback models automatically. Verifier: let the model perform its native self-checks, then add one external verifier for the final artifact, state change, or side effect rather than repeating internal prompt-level recheck instructions. Budget: set effort tier, token ceiling, runtime ceiling, spend threshold, and classifier fallback behavior before the run starts. Artifacts: keep the task brief, loaded context, tool trace, verifier result, and final output together so operators can inspect why a run succeeded or failed. Stop condition: the loop ends when the requested artifact or state change passes the verifier, or when the task hits its budget, permission, or confidence boundary and must escalate.

Take one long agent prompt and split every instruction into six buckets: task, reusable context, policy, tool access, verifier, and budget. Anything that stays in the prompt only because the old model needed hand-holding is a candidate to move into the runtime or delete outright.