The model is becoming the agent

The labs are absorbing orchestration — and selling it

Here's the tell. The same companies baking sub-agents into their models are also selling orchestration as a standalone product. OpenAI puts sub-agents inside GPT-5.6 and ships Codex, a cloud service whose manager decomposes work and dispatches it to parallel workers, each with its own context window. Anthropic gives Opus 4.6 native sub-agents and sells Managed Agents — composable APIs that handle the infrastructure, state, permissioning, and upgrade compatibility. If orchestration were really absorbed into the model, you wouldn't also need to buy it as managed infrastructure.

So what's being absorbed is the convenience layer — the happy-path glue you wired up to get a demo working. What's being sold is the part a model can't be: durable state, permissioning, retries, audit — the things that make an agent survive contact with production. The frameworks aren't dying. They're splitting in two.

Benchmark capability is not production reliability

The reason the hard layer survives is arithmetic. A model that absorbs the easy 80% still leaves the 20% where money and trust live — and that 20% compounds. An agent that chains 20 tool calls, each 95% reliable, finishes the whole job only about a third of the time (0.95^20 ≈ 36%). Multiple 2026 analyses of deployed agents tell the same story: leaderboard scores overstate real-world reliability, and success rates fall sharply as tasks lengthen and as systems move from demo to production.

A better single model raises the per-step number. It does not erase the compounding, the retries, the monitoring, or the question of who is accountable when step 14 of 20 quietly does the wrong thing. That work doesn't live in the model. It lives in the layer around it.

The argument this site exists to make

And then there's the failure mode no benchmark scores. On June 12, a US Commerce Department directive ordered Anthropic to suspend Fable 5 and Mythos 5 for any foreign national, inside or outside the United States — including the company's own non-citizen employees. Because the platform can't verify nationality in real time, the only compliant move was to take both models fully offline. The most capable models in Anthropic's history went dark in roughly an hour and a half, and stayed dark for weeks.

This is the counterweight to 'just use the one best model.' When your single great model can be gated by a government faster than your on-call can read the incident page, multi-vendor routing, graceful fallback, and lock-in resilience stop being architectural luxuries — they're the only thing between a policy letter and your product going down. Concentration is the risk; the layer around the model is where you manage it.

So the synthesis: agent frameworks aren't dying, they're bifurcating. The commodity convenience layer — decompose, fan out, run the steps — is being absorbed into the models, and will keep getting cheaper and better for free. The resilience, governance, and multi-vendor layer is becoming more valuable, not less — because models getting more powerful is exactly what makes depending on any single one more dangerous.

The takeaway