AI just hit three ceilings at once — cognitive, physical, and architectural

Three numbers landed today that should reshape what you're building this quarter.

First: BCG, in HBR, found that knowledge-worker productivity gains from AI reverse at the fourth simultaneous tool, and that the optimal dose is 7–10% of work hours — roughly 25–45 minutes a day. Past that, ActivTrak's behavioral data shows users spend 2× more time on email and 9% less on focused work. The study spans marketing, HR, ops, engineering, finance, and IT. It's self-reported and selection-biased. It's also the first credible quantification of something every operator has felt: there is a dose-response curve to AI tooling, and most teams are already overdosed.

Second: every frontier lab is now GA at 1M tokens — Gemini in early 2024, OpenAI on March 6, Anthropic on March 13 with Opus 4.6 hitting 78.3% on MRCR v2 and dropping the long-context surcharge. Two years of essentially zero growth in window size. The bottleneck isn't algorithmic. It's HBM and DRAM at inference sites, and semiconductor analysts converge on a 2–5 year ceiling. Sam Altman's 100× promise is hardware-disconnected. If your roadmap contains the phrase "when context grows to 10M, we simplify X," you should reclassify that line as a 2028+ horizon and stop waiting.

Third: OpenAI's Codex grew 5× in Q1 2026, and the growth driver wasn't the IDE extension — it was a standalone "mission control" app where developers run five parallel agent sessions at once. In a long technical interview, Codex lead Michael Bolin drew a sharper line than most of the industry has: the harness is the moat, not the model. Sandboxing, agent loop, tool surface, context management, training-inference format alignment. The single most impactful Codex improvement at the o3/o4-mini launch wasn't model scaling — it was making the training tool format match the production harness exactly.

These three findings rhyme. They all say: less is more, in a different layer of the stack.

The cognitive ceiling reprices your TAM

If you're selling an AI point solution into the enterprise, the BCG number is a problem. The pitch deck's bottom-up TAM almost certainly assumes unlimited per-seat adoption. The actual ceiling is one of three tool slots, used 10% of the workday. That's a fraction of the addressable surface most decks model.

The winners under this constraint are not the cleverest features. They are the consolidators — the products that absorb three tools' jobs into one interface, raising the user's effective ceiling rather than competing for the fourth slot. And the workflow-substitution plays — the systems that replace a job function entirely, where the BCG ceiling doesn't apply because no human is doing the tool-switching.

Meta's $600B capex commitment alongside ~15,800 layoffs is the market betting on substitution, not augmentation. Whether that's the right bet at that price is a different question. But the directional read is clear: capital is flowing to systems that eliminate workflows, not to assistants that add another tab to someone's browser.

The hardware ceiling makes retrieval a permanent discipline

For two years, "better RAG" has been treated as a temporary scaffold — something you'd rip out once context windows got big enough. They're not getting big enough. Build the retrieval layer like you build a database schema: with versioning, quality monitoring, eviction policies, and an owner.

Anthropic dropping the long-context surcharge while leading on MRCR v2 is the commoditization signal. Raw context is becoming table stakes. Quality at the ceiling — what you retrieve, how you compress, how you prioritize — is the new axis.

The IndexCache result (1.82× prefill, 75% of indexers removed) and the broader pattern of cross-layer KV reuse matter precisely because the window can't grow. Making the same 1M tokens cheaper and faster is the only lever left. If you run inference at any scale, that's where this quarter's optimization budget should go.

The harness is the moat — and the security boundary

Bolin's most useful framing is the security/safety split: security (sandboxing, access control, blast radius) lives in the harness. Safety (whether the agent should make this tool call at all) lives in the model. Fork the open-source Codex harness, swap in a non-OpenAI model, and the cage holds while the safety guarantees evaporate. This is elegant, and it's also the exact failure mode most agent frameworks haven't reckoned with.

If you're building model-agnostic agent infrastructure — which the open-source ecosystem largely is — the harness has to carry the full safety burden. Tool allowlists. Destructive-operation confirmations. Output validation. Trajectory mining (IBM's approach added 14.3pp on hard scenario goals, more than most model upgrades will deliver). Codex's "few powerful tools" philosophy — give the agent a terminal, not twelve specialized file APIs — works because shell commands are heavily represented in pretraining. In-distribution tool use is more reliable. The implication for your stack is testable: A/B a terminal-primary tool surface against your current specialized catalog and measure the failure rate.

NanoClaw hitting 22K stars in six weeks with a Docker Sandboxes integration is the other half of this story. Container-based agent isolation is converging into infrastructure quickly. If you're running agents that execute arbitrary code and you don't have a sandboxing strategy named in a design doc, that's the gap to close.

What to do this week

Three concrete moves, in order.

Count the AI tools your team actively switches between in a day. If it's four or more, consolidate before you ship the next one. The BCG ceiling is cheaply testable on your own team — instrument focused-work time and email volume, then cut a tool and re-measure in two sprints.

Audit your roadmap for any feature that assumes context windows past 1M. Reclassify those as 2028+ and redirect the engineering hours to retrieval quality, hierarchical summarization, and KV-cache optimization. Treat your RAG layer like infrastructure, not scaffolding.

Classify every guardrail in your agent stack as harness-side or model-side. Anything that lives only in the model is a guarantee you lose the moment a provider changes a default or someone forks your harness. Move the load-bearing controls into code you own.