1. AGENTS.md vs. skills: How to steer a coding agent — CircleCI

  • Why read: A practical guide to structuring prompts so coding agents actually follow them.
  • Summary: Stop dumping all your instructions into one file. Agent compliance drops as context length grows. Put universal rules like build commands and style guides in an `AGENTS.md` file. Put situational, complex workflows into modular skills that the agent can load only when needed. To know if this works, measure instruction adherence in your CI pipeline using test fixtures instead of guessing.
  • Read more

2. GLM-5.3: How Chinese labs keep stride with the frontier — interconnects.ai

  • Why read: How smaller labs match frontier American models using fast release cycles and heavy post-training.
  • Summary: Z.ai’s GLM-5.3 hits frontier coding performance with only ~750B parameters. Instead of relying entirely on pretraining, Chinese labs are scaling post-training and shipping updates in days rather than months. This fast iteration lets them climb benchmarks quickly using real-world feedback. A growing reinforcement learning data industry in China provides the training environments needed for these gains. As these highly capable, open-weight models spread, managing their safety at scale becomes a pressing issue.
  • Read more

3. Why model routing must be in the harness — X (formerly Twitter)

  • Why read: Why you should move LLM routing logic into the agent harness to cut costs and latency.
  • Summary: Long-running agents are expensive. To cut costs, enterprises need dynamic model routing. Doing this at the gateway layer fails because gateways cannot see the agent's task history, cache state, or tool outputs. If you route inside the harness instead, the agent can choose cheap models for simple exploration and frontier models for hard reasoning. This approach cuts costs by up to 58% and lowers latency by taking advantage of warm caches and task context.
  • Read more

4. Revised rules of engineering leadership. — lethain.com

  • Why read: How engineering management needs to change now that AI tools let individuals do the work of teams.
  • Summary: Single engineers can now execute massive migrations. Because work happens at machine speed, individual judgment matters more than ever; small mistakes compound fast. AI makes writing the first draft of code cheap, but the cost of shipping depends entirely on the quality of your testing and CI/CD harness. Leaders should automate routine tasks like initial code reviews and issue triage. Keep long-standing teams together to preserve domain knowledge. To keep up with the new pace, executives must make decisions quickly and stick to them.
  • Read more

5. Cursor + SpaceXAI: the fastest iterating team wins — X (formerly Twitter)

  • Why read: Breaking down SpaceXAI's $60 billion acquisition of Cursor and the future of coding environments.
  • Summary: Cursor won developer mindshare by focusing on product experience and model flexibility rather than building foundation models. When agentic models threatened IDEs, Cursor used its proprietary data to build custom Composer models. They eventually hit a wall because they lacked the raw compute to compete with OpenAI and Anthropic. The SpaceXAI acquisition solves this, combining Musk’s gigawatt-scale data centers with Cursor’s fast product iteration. Expect specialized, tightly integrated AI coding environments to become the default for enterprise development.
  • Read more

6. Understanding is the new bottleneck — geoffreylitt.com

  • Why read: As agents write more of our code, human understanding replaces code generation as the main bottleneck.
  • Summary: AI can write the code, but humans still need to understand it to verify correctness, steer product direction, and avoid cognitive debt. Raw diffs are no longer enough. Engineers should use AI to generate interactive explainers, custom debuggers, and quizzes to test their own comprehension of the codebase. The next challenge is building shared team environments to collaboratively read and reason about AI-generated code.
  • Read more

7. Verifiable Since 1494 — X (formerly Twitter)

  • Why read: Why accounting—a field with built-in mathematical verification—will likely produce the first trillion-dollar AI-native services firm.
  • Summary: Professional services break down into work that is graded (code, accounting) and work that is argued (law). AI coding took off because compilers provide free, instant grading. Accounting has a similar built-in eval suite: the double-entry ledger, invented in 1494. Debits must equal credits and tie to bank statements, so AI-generated financials can be verified automatically without human debate. As AI handles this verification near zero cost, the industry will move away from delayed monthly closes and random sampling.
  • Read more

8. Honestly, Who Buys SOTA? — X (formerly Twitter)

  • Why read: Why usage is shifting away from frontier models toward cheaper, good-enough alternatives.
  • Summary: Most people are not paying for state-of-the-art models. On platforms like OpenRouter, 84% of processed tokens use highly efficient models that offer roughly 80% of frontier performance for about 2.5% of the price. Startups and enterprises are optimizing for cost on routine workloads. Frontier models remain necessary for complex tasks like architecture and security, but they may struggle to capture enough volume to pay for their nine-figure training runs. Spending is consolidating around smaller, open-weight models.
  • Read more

9. Training Frontier Review Table Models with Applied Compute — X (formerly Twitter)

  • Why read: A case study showing how a specialized model paired with agentic retrieval beat general-purpose LLMs in legal document analysis.
  • Summary: Harvey worked with Applied Compute to train a custom model for parsing large legal data sets. They used a synthetic dataset and a reward function that penalizes fake citations to beat the accuracy of Claude 5 and GPT-5.6-Sol. They also moved from single-turn semantic search to multi-turn agentic retrieval, cutting input token usage by 50.3% without hurting quality. The project proves domain-specific post-training and smart retrieval can outperform general frontier models while lowering costs.
  • Read more

10. The AI Engineering Skills Map — X (formerly Twitter)

  • Why read: Andrew Ng maps out four skills developers need to become AI engineers.
  • Summary: AI outputs are unpredictable, requiring a different approach than deterministic code. Developers need four skills to adapt. First, deploying AI apps, which requires knowing LLMs, RAG, and statistical evaluation. Second, solid software engineering fundamentals to make architecture choices and prompt agents with precise language. Third, mastering coding agents—managing their context, orchestrating them, and knowing when to step in. Fourth, product sense; engineers now have to write the specs and define the goals the agents will follow.
  • Read more

11. Deep Dive: The Next Trillion-Dollar Futures Market — Substack

  • Why read: How the new CME Group compute futures contracts aim to stabilize the volatile AI infrastructure market.
  • Summary: AI capital expenditures are projected to hit $765B in 2026, passing oil and gas. To manage the resulting financial risk, the CME Group is launching compute futures so companies can hedge against GPU rental price swings and hardware depreciation. Currently, the market lacks stabilization; buyers and builders face huge risks from new chip releases or sudden demand spikes. The biggest hurdle for these futures is that identical GPUs perform differently depending on the cloud provider. If the market standardizes, compute futures could become a massive new asset class.
  • Read more

12. Eighteen Models, One Data Room — Substack

  • Why read: Testing 18 LLMs in a synthetic corporate data room to see which models can actually handle due diligence.
  • Summary: Standard benchmarks do not reflect real corporate diligence work. A synthetic data room tested 18 models on spotting deal-breakers and suggesting fixes. Frontier models (Fable 5, Opus 5, and Sol 5.6) nailed the basic recall tasks but varied in judgment. Fable 5 proved the most reliable for high-stakes decisions. Sol 5.6 and Meta's Muse handled volume triage well at a lower cost. Weaker models missed too much to be trusted. While frontier models cost more, paying the premium is cheaper than missing a fatal flaw in a major deal.
  • Read more

13. Clouded Judgement 8.14.26 - Useful Life — Substack

  • Why read: Why older GPUs are holding their value longer than the expected 4-5 year lifespan.
  • Summary: The industry assumed GPUs would depreciate quickly. Coreweave data shows older chips like the 2020 A100 are being re-contracted out to 2029 at solid prices. While frontier models need the newest hardware, a massive tail of simple inference tasks runs fine on older chips. This extends the useful life of the hardware and generates pure profit for neoclouds that have already paid off their initial debt on the chips. Software that routes workloads to the correct chip tier will become highly valuable.
  • Read more

14. Rules of thumb for reorgs/org design: — X (formerly Twitter)

  • Why read: Heuristics for when and how to restructure engineering teams.
  • Summary: Early startups need cross-functional teams. Scaling companies need to separate teams by the type of work they do, such as product versus infrastructure. If a strong team slows down, they are usually juggling too many types of work and need to be split. Accept that communication degrades across team boundaries; draw your org chart where you can afford that friction. For massive initiatives, pull people into a temporary, dedicated team. Reorgs are painful in the short term but necessary to fix structural drag.
  • Read more

15. AI Is Becoming a Compiler for Intelligence — Substack

  • Why read: Why the best use of an expensive frontier model is writing cheap scripts to replace itself.
  • Summary: Do not use a massive model to run every step of a repetitive task. Have the model figure out the problem once and write a simple script, rule, or smaller model to handle it going forward. This converts an ongoing inference cost into a permanent, nearly free asset. As AI turns complex tasks into simple automated loops, it cannibalizes its own inference volume. The frontier models will move on to harder edge cases, leaving a trail of cheap automation behind them.
  • Read more

Themes from yesterday

  • Efficiency wins: Volume is moving from state-of-the-art frontier models to cheaper, smaller, or specialized models to keep inference costs down.
  • AI writing automation: Instead of running expensive models on a loop, engineers are using them to generate simple scripts and rules to automate tasks permanently.
  • Compute holds its value: Older GPUs are lasting longer than expected, and financial markets are launching futures contracts to hedge against hardware price swings.
  • Routing moves to the harness: Routing LLM requests dynamically within the agent harness—rather than at the gateway—saves money by matching the model to the exact difficulty of the sub-task.