Episode #10

This run of "Agentic by Default" examines the hidden costs and complex measurements that define agentic AI systems. We ask if the gains from memory layers truly outweigh their expense and if the promise of proactive agents can be delivered responsibly. We also scrutinize how we quantify agent performance and its impact on human work. The central tension is between the perceived efficiency of agents and the less-discussed trade-offs and methodological challenges involved in their development and deployment. This episode delves into the real calculus behind the agentic future.

The "Memory Tax" article asks whether a designed memory layer still pays off once a model can search its raw history directly. Its tests suggest a designed memory layer is not always the optimal path. For instance, Mem0 beat full context only when paired with the weakest model in one test conversation, and building its memory cost more than a whole full-context rollout. This raises direct questions about the architectural benefits versus the overhead.

Is the memory layer always capped by the model that writes it, preventing a stronger reader from recovering what a weaker writer left out?

Yes, the article highlights this limitation clearly. A designed memory layer’s effectiveness is fundamentally limited by the intelligence of the agent that creates it, regardless of the retrieval model's strength. This suggests a systemic weakness in relying on less capable agents to build foundational components for stronger ones.

Anthropic has opened a waitlist for Claude Duet, a proactive agent, as a research preview for US Pro and Max subscribers. Duet is designed to send a morning brief and check connected apps every few hours, messaging you only when something needs attention. These background checks are significant because they count against your plan's usage limits, which directly shapes the agent's design and user experience. The owner of this publication uses Anthropic's products.

Duet shows what it will do and waits for approval before taking actions like sending or paying; are any actions barred even with approval?

Yes, Anthropic states that even with approval, trading securities and moving money between accounts are explicitly off-limits for Duet. This indicates Anthropic is drawing clear boundaries around financial agency. However, their pages leave open questions about data retention, the use of Duet data for training, and pricing beyond the Pro and Max tiers.

Anthropic's study claims agent hours worked outnumber human hours 6 to 1, but this figure compares agent runtime with an assumed 40-hour human week. It counts how much work was handed to agents, not how much was actually done. The more critical numbers concern the human impact: the share of junior technical hires fell from 27% to 9%, and meta-work rose from 10% to 22% of interactive sessions, with no cause named in the study. The owner of this publication uses Anthropic's products.

The study uses automated Claude graders. Is there a reported accuracy figure for them?

No, the study from Anthropic does not report an accuracy figure for its automated Claude graders. This detail, alongside the fact that the study is Anthropic measuring itself, raises questions about the overall reliability and interpretability of its findings. When the tools of measurement are not themselves measured, the resulting data lacks crucial context for systemic evaluation.

The run presents a central tension between the perceived benefits and efficiency of agentic systems and the complex, often unstated, costs and measurement challenges involved.

The "6:1" figure is a perfect example of this. It presents a headline number without clarifying that it counts hours *handed over*, not *work completed*. This kind of metric can easily mislead about true productivity gains.

But the Memory Tax article shows that even when we attempt direct cost comparison, a "costly fix" might emerge where a designed memory layer costs more than a full context rollout. This points to inherent complexities in system design.

And Duet's usage limits, counting background checks against a user's plan, show that these costs aren't abstract; they directly influence how a proactive agent can operate. Cost structures shape the agent's actual "proactiveness."

Exactly. The human impact is obscured. Falling junior hires and rising meta-work aren't just statistics; they point to shifts in the nature of human work alongside agents, which the 6:1 figure utterly bypasses.

From an architectural perspective, if a memory layer is capped by the weaker model that wrote it, then building stronger agentic systems requires more than just adding layers. It demands a holistic improvement in component capabilities.

Users also face a control versus cost dilemma. Duet offers proactive assistance, but the checks cost money, and significant actions like trading securities are always off-limits. The perceived convenience comes with clear boundaries.

We need to push past the surface claims. If the automated graders in Anthropic's own study have no reported accuracy, how much confidence can we place in any of their self-measured metrics?

This is about verifiable system performance. If a Pi agent achieves 99% accuracy at less than half the cost, but only on one conversation and partly due to photo captions agents could read, its generalizability for broad application remains an open question.

So we're left with the immediate user-facing implications: agents are expensive to develop, expensive to run, and their promised efficiencies often come with caveats about control and unmeasured performance.

This run left several gaps in our understanding of agentic systems. For the Memory Tax, it is unclear if the Pi agent's cost-efficiency holds beyond the single conversation and single run per cell reported. The paper tested nothing on its own. Regarding Duet, critical details like data retention policies, whether Duet data is used for training, and pricing beyond Pro and Max remain undisclosed by Anthropic. Finally, for the 6:1 measures, the study names no cause for the observed shifts in human roles, leaving us without insight into the driving factors behind these significant changes.

This episode clarifies that the journey toward truly agentic AI is paved with intricate trade-offs. We cannot simply accept claims of efficiency without scrutinizing the underlying costs, measurement methodologies, and human impacts. Understanding what we pay, what we control, and how we measure is fundamental. The rings of this run highlight the ongoing challenge of assessing agentic value with precision and transparency.