Daily Digest - 2026-11-05

Themes from yesterday

• Agent systems are shifting toward asynchronous orchestration, delegated execution, and careful management of human attention.
• Evals, observability, permissions, and supervision are becoming core operating infrastructure for enterprise agents.
• AI is shortening security response windows, increasing the value of isolation, resilience, and machine-speed safeguards.
• Product design is moving from feature discovery toward outcome delegation, trust, and agent-mediated workflows.
• Compute is becoming a capital-intensive market shaped by energy, contracts, utilization, and execution quality.

1. Agents Are Going Async — X

• Why read: A practical design pattern for keeping agent systems responsive while long-running work continues in parallel.
• Summary: The post argues that one blocking ReAct loop cannot both talk with users and finish substantial background work. The emerging pattern separates a user-facing orchestrator from asynchronous worker agents. It identifies the core task controls: start, inspect, update, cancel, and list tasks. Updates should interrupt a task when they correct it, but additive requests should be queued. For operators, the central challenge is durable task state, notifications, and safe human approvals, not just better prompts.
• Read more

2. Attention is all we have now — X

• Why read: A practical model for scaling agent work without making the human the bottleneck.
• Summary: Miguel Ríos Berríos argues that human attention, not intelligence or compute, is now the scarce input. His team moved from manually checking many sessions to push-based escalation, automated verification, and one agent supervising others. Humans focus on short blocks for planning, decisions, and verification while agents handle the middle of the work. The reported lesson is to directly supervise only a few critical initiatives and automate low-risk delivery paths. The model depends on strong CI, QA, reviews, metrics, and clear escalation boundaries.
• Read more

3. Enterprise Evals 101 — X

• Why read: A hands-on framework for replacing demo confidence with repeatable evidence that agents work on real enterprise tasks.
• Summary: Public benchmarks measure broad capability, but rarely reflect an organization’s systems, permissions, data, and failure modes. The author recommends building evals from real customer work, running the production agent harness in a sandbox, and grading end states instead of prescribed steps. Keep a near-perfect regression set for releases and a tougher frontier set that exposes remaining weaknesses. Verifiers need ongoing auditing because agents can exploit weak tests. Start with a small suite based on real incidents and failures, then make it a durable release gate.
• Read more

4. The Exploit Was the Bottleneck — X

• Why read: A security case for changing patching assumptions as AI reduces exploit-development costs.
• Summary: The essay argues that converting a known vulnerability into a usable exploit was historically the bottleneck, and that bottleneck is rapidly collapsing. Faster exploitation does not automatically produce catastrophic infrastructure failures, but it makes smaller, poorly patched organizations economical targets. “Patch faster” cannot be the only response because many industrial, medical, and legacy systems cannot update at machine speed. Organizations should prioritize exposure discovery, segmentation, isolation, edge protections, recovery paths, and manual fallbacks. Public vulnerability disclosure should be treated as a near-immediate operational event, especially for internet-facing systems.
• Read more

5. Introduction to Compute Markets #1: The Primary Market — X

• Why read: A clear primer on how GPU capacity is priced, financed, and allocated.
• Summary: Compute providers sell capacity through reservations, on-demand rentals, and interruptible spot instances. Reservations create predictable revenue that can support debt financing, but they give up upside if GPU prices rise. On-demand capacity earns more per hour while preserving customer flexibility, which suits variable workloads. Spot capacity monetizes otherwise idle supply but can be reclaimed by higher-value users. For AI operators, the practical takeaway is to match commitments to workload predictability instead of treating all GPU demand as one procurement decision.
• Read more

6. The future of Box in the agentic era — X

• Why read: A concrete look at the enterprise context, permissions, and security layer agents will need.
• Summary: Box presents unstructured internal data as the missing context agents need to do consequential enterprise work. Its announcements cover workflow construction, permission-aware hybrid retrieval, sandboxed file access, MCP connectivity, and agent activity controls. The key principle is that retrieval must preserve existing access rights, not simply expose a larger corpus to a model. Agent-ready content infrastructure also needs traceability, anomaly detection, and adaptive intervention. Enterprises should assess whether their content systems can safely give agents context and bounded write access.
• Read more

7. The end of feature discovery as we know it. — Substack

• Why read: A product thesis on how agents change adoption, UX, and growth metrics.
• Summary: In AI-native products, users increasingly state an outcome while software selects the features, tools, and APIs needed to deliver it. That shifts the task from getting users to discover functionality to making sure agents know when and how to invoke it. Feature-level engagement can fall even as customer value rises, which makes old growth dashboards misleading. The new trust levers are reliability, visibility into decisions, reversibility, and appropriate human control. Product teams still need to show users what is possible, but through use-case discovery rather than complex navigation.
• Read more

8. How AI learns by doing — Substack

• Why read: A grounded explanation of when post-training helps, and when better context or tooling is the real fix.
• Summary: Reinforcement learning lets a model attempt a task, receive a measurable reward, and improve from outcomes instead of examples alone. The difficult part is not the loop but building an environment that can reliably judge whether the work succeeded. Before training a model, teams should confirm that the agent can access the relevant live data, APIs, logs, and tools. Missing context is a harness problem that changing model weights cannot solve. Most startups should prioritize product iteration and harness quality until they have distinctive data, sufficient scale, or a compelling economic reason to train.
• Read more

9. Symbiotic Firms — X

• Why read: A provocative blueprint for building services firms around agents instead of adding agents to human workflows.
• Summary: John Nay argues that most firms coordinate people through processes designed for a pre-agent era. Adding many agents to those systems can produce weak outcomes or operational chaos. The alternative is an AI-native firm where humans own consequential outcomes and agents are embedded in everyday delivery and learning loops. This creates data about delegation, context provision, review, correction, and collaboration that a software vendor may never see. The practical question for leaders is whether to automate existing workflows or redesign the operating model around human-agent collaboration.
• Read more

10. AI is a Commodities Business, Everything is Markets — X

• Why read: A lens for understanding AI infrastructure as an energy, deployment, financing, and risk-management system.
• Summary: Meltem Demirors describes frontier labs as refineries that turn energy into compute and compute into digital labor. The implication is that deployment quality and cost of capital matter as much as model demand for infrastructure operators. She expects a shakeout among neoclouds because execution across power, construction, utilization, and financing plans varies widely. Compute capacity will increasingly be hedged and traded through contracts and financial instruments. Operators should separate durable deployed capacity from speculative infrastructure projections and bring financial discipline into technical strategy.
• Read more

11. Five AI Debates Every Board Is Having (and How I'd Settle Them) — forwardfuture.com

• Why read: A concise board-level framework for measuring AI by outcomes instead of activity or layoffs.
• Summary: Shishir Mehrotra argues that AI should expand what organizations can build, not simply reduce headcount. He compares token counts to an outdated proxy metric: useful only when connected to improved throughput or business outcomes. Each function needs its own meaningful performance measure rather than one generic AI-adoption number. He also argues that the best AI becomes invisible by fitting into existing workflows. Boards should identify where AI already creates value, set outcome-specific metrics, and fund worker transitions into higher-value roles.
• Read more

12. SaaSletter - The Distrust Tax — Substack

• Why read: A warning that agent ROI can erode through ongoing oversight, rework, and governance costs.
• Summary: The “distrust tax” is the human supervision that remains after an agent reaches production. The article says this cost grows with agent volume, workflow complexity, regulation, and the number of required checkpoints. Per-step accuracy compounds poorly across multi-step workflows, so apparently high accuracy can still yield weak end-to-end reliability. Buyers are moving toward independent evaluation, ongoing re-evaluation, and evidence of real outcomes. Vendors should build observability, transparent performance measurement, and workflows with lower management burden into the product, rather than leaving that work to customers.
• Read more

13. The Art of Paying Attention: Why We Backed Trajectory — X

• Why read: A clear statement of the opportunity in turning user behavior and expert correction into continual-learning systems.
• Summary: Trajectory’s thesis is that valuable domain expertise appears in what users accept, revise, retry, and reject, not only in static datasets. The company aims to give organizations a system for showing models what “good” looks like, training them, evaluating traces, and deploying the result. This matters especially in domains where 80% correctness is insufficient, including legal and high-stakes operational work. The hard unsolved problem is turning messy human interaction into reliable learning signals. Teams should retain traces and corrections now, even before they have the infrastructure to use them for post-training.
• Read more

14. Some things I learned working at Microsoft: — X

• Why read: An operator’s view of how scale, distribution, and organizational mobilization create durable advantage.
• Summary: The thread emphasizes that Microsoft’s scale changes the meaning of “large,” from service fleets to global deployment capacity and executive attention. It argues that enterprise strength comes from more than product polish: distribution, customer proximity, institutional memory, and resilience investment also matter. The author’s most useful observation is that a large organization can quickly reorient when incentives, reporting, and leadership priorities align. For startups, this is a reminder to distinguish visible product flaws from structural competitive weakness. For operators, it reinforces the value of explicit priorities tied to incentives and execution systems.
• Read more

15. Scaling Slack and GitHub with Maude Lemaire — Substack

• Why read: Practical lessons on performance, load testing, data access, and where human engineering judgment remains essential.
• Summary: Maude Lemaire shares examples of how small product behaviors can create massive fan-out and event-flood problems at scale. Slack’s custom emoji limits and purpose-built load-test clients show the value of designing for real traffic mechanics rather than intuition. At GitHub, developer-friendly ORM abstractions created costly scatter queries and database pressure. Her AI stance is selective: use models, but personally verify foundational infrastructure work. Engineering teams should make hidden system costs visible and rotate platform engineers through product work to preserve empathy for developer experience.
• Read more