The Signal — Sam Altman quietly shipping Codex in mobile; @svpino declaring subagent-pilled workflow as largest 2-week shift; Computer Use getting major upgrade. Agentic composition beating monolithic models as the actual competitive moat.
Consensus: Bullish | Conviction: High
What's Moving
- Codex in ChatGPT mobile — OpenAI pushing code execution and task completion into consumer product; signals agents are moving from demos to monetizable UX (via @sama)
- Subagent architecture standardizing — Each subagent gets isolated context, model selection, tool configuration; pattern emerging as best practice for real agentic work (via @svpino)
- Computer Use upgrades rolling — New implementations focused on stability and cost—finicky agents being replaced by production-grade alternatives (via @svpino)
- Local inference hardware commoditizing — Sub-$1K mini computers (GEEKOM A9 Max) becoming viable for agentic workloads; hardware no longer the bottleneck (via @svpino)
- Open-source models proving sufficient for agents — DeepSeek Flash/Pro "just as good" at 20x lower cost; agentic loops not requiring frontier models (via @bindureddy historical)
Blind Spot — Consensus assumes bigger models = better agents. Reality: @svpino's shift to subagents suggests architecture beats parameters. The real unlock is orchestration—parallel execution, context isolation, tool specialization. This flips the competitive game from "train a bigger foundation model" to "who builds the best agentic router and task decomposition." Open-source and mid-tier models win here. Also: nobody discussing cost-per-completed-task vs. cost-per-token. That metric matters more for agents than raw inference speed.
One Actionable Idea — Track which framework/platform is becoming the "Rails of agentic composition"—the one practitioners default to for multi-agent orchestration. That company owns the next layer up from model providers.
Sources: @sama (Codex momentum), @svpino (subagent pattern + local inference), @bindureddy (open-source viability for agents)