1. How Kenn is doing Agentic Engineering in August 2026 — wesmckinney.com

  • Why read: A concrete workflow for using coding agents while keeping humans responsible for quality.
  • Summary: Kenn separates design, specification, implementation, verification, and merge ownership instead of treating agent output as production-ready. The workflow uses independent reviews, small implementation slices, frequent commits, and ongoing adversarial checks. Humans own the decisions and outcomes; agents speed up typing and testing. Keep specs and plans as temporary execution artifacts, then move durable knowledge into living architecture documentation.
  • Read more

2. Human Authorization in Agentic Workflows — Tempo

  • Why read: Shows how to prove that a human approved a sensitive agent action.
  • Summary: Authentication cannot tell the difference between an employee’s intended action, malware, and an agent misusing valid credentials. Tempo binds approval to a Touch ID event and Secure Enclave signature, with expiry and replay protection. The proof can be used in GitHub reviews, Kubernetes actions, and deployment workflows. For high-impact actions, tie human presence to the exact changeset or request rather than a generic approval token.
  • Read more

3. Right-Sizing Your Intelligence Spend — Jaya Gupta

  • Why read: Makes the case for routing work by the capability it actually needs.
  • Summary: Most enterprise tasks depend on context and execution, not frontier-level discovery. Once a model is capable enough, more reasoning can add cost, latency, and complexity without improving the result. The piece recommends routing work across frontier, open-weight, local, deterministic, and human systems under enterprise control. Measure intelligence cost per verified outcome, then turn repeated reasoning into cheaper workflows or rules.
  • Read more

4. Building software factories (with no slop) — David

  • Why read: Reframes AI code review around interpretation and verification rather than reading every line.
  • Summary: Agent output is already exceeding the capacity of human reviewers, so adding more reviewers will not scale. A better approach is to build independently verifiable components with clear seams, observable behavior, and explicit invariants. Agents should record the assumptions they make when the specification is silent. Review consequential decisions and behavioral evidence, ideally through an independent auditor rather than the original implementer.
  • Read more

5. Debate Training Reduces Reward Hacking in RLAIF — GDM Alignment Research Blog

  • Why read: Shows how adversarial critique can improve model-generated training signals when ground truth is unavailable.
  • Summary: Training against an LLM judge can raise the judge’s score while actual task accuracy declines. In the reported experiments, adding a second model to critique the answer recovered about 45% of the gap to ground-truth-supervised training. This is relevant to fuzzy criteria such as maintainability, research quality, and instruction-following. The method is incomplete: critics can also manipulate the judge, so the protocol still needs careful design and auditing.
  • Read more

6. Enterprises with AI context layers report agent failures at more than twice the rate of those without one — VentureBeat

  • Why read: A reminder that better instrumentation can make reported failures rise before they fall.
  • Summary: Enterprises with governed context layers report more recurring confident-but-wrong answers because they can trace errors to stale data and conflicting definitions. Retrieval cannot resolve a semantic conflict when two systems use the same term differently. A clean incident record may therefore indicate weak instrumentation rather than healthy agents. Establish shared definitions, provenance, evaluation, and failure tracing before treating low reported failure rates as success.
  • Read more

7. What every CEO needs to know about AI governance — Bessemer

  • Why read: A board-level checklist for governing data, models, and high-impact AI decisions.
  • Summary: Cybersecurity protects access to data; AI governance must also address harmful influence from poisoned, biased, or poorly sourced inputs. The proposed chain of custody covers origin, handling, tamper evidence, reproducibility, access logs, and preservation of model artifacts. Once data has shaped a trained model, tracing its provenance becomes much harder. Assign an accountable AI leader and make provenance requirements contractual for data and model vendors.
  • Read more

8. The Agent Harness Should Be Open — TrueFoundry

  • Why read: Shows how runtime design affects both agent reliability and operating cost.
  • Summary: The harness controls context compaction, tool-call frequency, state, sandboxing, delegation, and model portability. In the reported benchmark, a leaner harness achieved similar solve rates with fewer tokens than a managed-agent alternative. Treat the benchmark as vendor evidence, but take the broader point seriously: orchestration choices compound over long runs. Keep models, tool providers, and runtime components replaceable so cost and quality can change independently.
  • Read more

9. 2026: The Year of Dataflow — David Dong

  • Why read: Offers a systems view of why moving data can matter more than doing arithmetic.
  • Summary: Modern AI systems increasingly optimize data paths between SRAM, HBM, operators, chips, and racks. FlashAttention shows the trade-off: recomputing can cost less than writing and rereading large intermediate results from memory. The piece presents TPU, SambaNova, Groq, and Cerebras as different attempts to reduce movement and scheduling costs. For product builders, inference cost and latency increasingly depend on placement, cache locality, and scheduling as well as model choice.
  • Read more

10. Computers Compared — Fly.io

  • Why read: Provides a practical way to assess whether an agent environment can support ongoing operational work.
  • Summary: Agents need durable computers rather than short-lived sandboxes: persistent files, running processes, stable addresses, and securely delegated credentials. Those capabilities support webhooks, OAuth callbacks, long-running services, and stateful work without constant reconstruction. The article also distinguishes secret injection from scoped identity brokering, which limits damage when an agent makes a bad but authorized request. Use the rubric to separate platforms built for real operations from those designed for isolated jobs.
  • Read more

11. How can LLM RL Work Despite Information-Theoretic Inefficiency — Beren Millidge

  • Why read: Explains why sparse rewards can work well for targeted tasks.
  • Summary: Reinforcement learning looks information-poor because one reward may supervise an extremely long rollout. Its advantage is that the signal is aimed directly at the task, while next-token training contains much more information that may be irrelevant. In that sense, RL can have a higher task-specific signal-to-noise ratio despite using fewer bits. The practical lesson is to improve environments, verifiers, and outcome signals instead of assuming more training text is always the answer.
  • Read more

12. Why LLMs can’t make your code simpler — answer.ai

  • Why read: Explains why simplicity depends on system knowledge, not syntax alone.
  • Summary: Drawing on Peter Naur, the piece argues that the real program is the team’s working theory of the domain, trade-offs, users, and system behavior. Code and documentation only represent parts of that theory. Models tend to add complexity when they lack sustained experience with the system and its constraints. Keep humans close to domain learning, reduce scope deliberately, and treat agent-generated code as a proposal that still needs to earn its claim of simplicity.
  • Read more

13. Operational Data Rights - What are they and how valuable are yours? — Ryan Lock

  • Why read: Offers an early framework for assessing operational traces as an AI-era asset.
  • Summary: Connected records of tasks, communications, transactions, and outcomes reveal how work actually gets done. The essay points to completeness, extractability, legal title, exclusivity, and decay as factors in their value. Verify the market claims independently, but take the strategic question seriously. Inventory operational data, clarify rights and consent, and decide whether it is a product advantage, a licensable asset, or something better kept private.
  • Read more

14. The rise of the Executive Operator. — Musings from a 2x Chief of Staff

  • Why read: Describes a role built to turn leadership decisions into repeatable company-wide execution.
  • Summary: The Executive Operator owns planning cadence, dependencies, follow-through, decision records, and organizational clarity. Unlike a Chief of Staff focused mainly on one principal, this role supports the company’s operating system. The distinction matters as AI makes it easy to produce documents that look like alignment without creating accountability. If a Chief of Staff is handling both executive support and company-wide execution infrastructure, those may be separate jobs.
  • Read more

15. [AINews] Memory prices up 500% in 12 months — AINews

  • Why read: Connects memory constraints, model progress, inference economics, and evaluation practice across the AI stack.
  • Summary: Memory supply and pricing are becoming binding constraints alongside compute and power. The roundup also covers capable smaller open models, post-training gains, faster inference, and continuous evaluation on production traces. Together, these trends shift competition away from raw model access toward harness design, systems optimization, and reliable feedback loops. Treat memory, routing, and observability as core roadmap constraints.
  • Read more

Themes from yesterday

  • Agentic engineering is becoming a practice built on clear specifications, verifiable boundaries, independent review, and durable operational knowledge.
  • The harness, context layer, and execution environment increasingly shape agent cost, reliability, and security.
  • Route intelligence by workload requirements and optimize for verified outcomes rather than token volume.
  • Governance now covers provenance, authorization, traceability, and named ownership alongside access control.
  • AI infrastructure constraints are spreading from GPUs to memory, data movement, persistent compute, and rights over operational data.