Three new cards from two operators across engineering and AI-native lanes. The scan pulled from a quiet Sunday and Monday in operator publishing. One convergence worth noting: all three cards are ultimately about design constraints that are architectural, not scaling problems. You cannot solve accountability by adding more agents. You cannot solve agent forgetting by expanding context. You cannot serve every AI workload cheaply from one general-purpose inference layer.
Theme 1, Accountability is the last function machines cannot inherit from humans
Simon Willison published a July 12 essay on why AI agents should never hold the DRI role on any project. AI agents cannot hold the DRI role because accountability requires the capacity to own consequences, which machines lack captures the argument: the DRI role requires a party who can be held accountable for outcomes, not just a party who executes. Machines make decisions. They cannot own consequences. Willison traces the principle to a 1979 IBM training slide stating that "a computer can never be held accountable, therefore a computer must never make a management decision." The observation predates the current AI wave by decades, which is part of what makes it durable.
The mechanism is structural, not technological. Accountability involves social, legal, and reputational capacity to answer after the fact. A machine that produces a wrong output cannot be questioned, demoted, or held responsible in the way a person can. Assigning DRI to an agent does not distribute accountability; it eliminates it. The practical implication is that agent-heavy workflows still require named human owners at every decision point that matters externally.
- Simon Willison, AI agents cannot hold the DRI role because accountability requires the capacity to own consequences, which machines lack. AI agents should never be assigned as the DRI for a project; accountability is inherently human.
Theme 2, Architecture beats scale: the design choices that actually determine whether AI systems stay reliable
Two Tunguz posts from July 8 and July 10 surfaced a common structural argument: the hard limits in AI systems are not raw scale or model capability but design choices upstream of execution.
The July 8 post, captured as Agent forgetting is a memory architecture problem, not a context size problem, and the fix is selective skill retrieval before every task, makes this concrete for agent memory. Tunguz describes a preflight check borrowed from aviation: before executing any task, the agent inspects its skills library, matches the query to relevant skills via intent matching, and loads only those into context. The claim cuts against a common assumption. When agents forget mid-task, the instinct is to buy a bigger context window. Tunguz argues the real fix is not size but selectivity: "Context size is not the ceiling. Memory architecture is." A background watchdog logs overnight, hardens successful patterns into deterministic code, and keeps the system improving without retraining.
The July 10 post, captured as AI inference infrastructure is fragmenting into specialized layers the way databases fragmented into OLTP, OLAP, vector, and streaming systems, zooms out to the infrastructure layer. Tunguz observes that AI inference is now specializing by workload: video, batch, local, agentic, and real-time each carry irreconcilable trade-offs in latency, cost, and throughput. The parallel he draws is to databases: OLTP, OLAP, vector, and streaming emerged because no one tool could optimize across all those axes simultaneously. Inference infrastructure is at the same inflection point. The implication for teams building on AI is that "which model" is the wrong frame. The right frame is "which model for which workload."
- Tomasz Tunguz, Agent forgetting is a memory architecture problem, not a context size problem, and the fix is selective skill retrieval before every task. Agent memory failures are an architecture problem; selective skill retrieval at task start outperforms context expansion.
- Tomasz Tunguz, AI inference infrastructure is fragmenting into specialized layers the way databases fragmented into OLTP, OLAP, vector, and streaming systems. AI inference infrastructure is fragmenting into specialized layers by workload type, repeating the database fragmentation pattern.