1. How Foundational Models Became Superhuman in Bash — X (formerly Twitter)

  • Why read: Shows why AI models use Bash to run multi-step command-line tasks faster than humans.
  • Summary: Coding agents are adopting Bash as a routing layer instead of relying on custom, single-purpose tools. By writing throwaway shell scripts on the fly, models can handle multi-file edits, track down flaky tests, and parse logs. This fits the "Bitter Lesson"—general methods paired with compute win out over hand-crafted shortcuts. If you build agent environments, the takeaway is to stop writing micro-tools. Give the agent a clean shell environment with basic safeguards like truncated outputs and background process handling. You get a smaller interface but give the agent way more room to work.
  • Read more

2. How We Rebuilt Playbook Review as a Multi-Agent System — X (formerly Twitter)

  • Why read: A look at how Harvey moved from a simple prompt pipeline to a multi-agent setup for legal review.
  • Summary: Checking contracts against legal playbooks involves conditional logic and cross-document dependencies that break single-prompt pipelines. Harvey fixed this by using a lead agent to dispatch worker agents, letting them evaluate separate rules at the same time on different branches of the document. The system reconciles conflicting edits, shares context to keep track of the deal's status, and saves state for follow-up questions. To make it work in production, they added timeouts, cached prompts, streamed results, and routed tasks to different models to balance cost and speed. The result is a system that can reliably review contracts in the background.
  • Read more

3. A2A Is the Task — torresmateo.com

  • Why read: Compares how the Agent-to-Agent (A2A) protocol and the Model Context Protocol (MCP) handle stateful tasks.
  • Summary: MCP tasks work like temporary receipts for one-off requests. A2A tasks act as persistent threads. A2A keeps a shared history of messages from both client and server and delivers artifacts as work gets done. It treats waiting for user input or approval as standard task states, rather than hacks. This makes it easier for remote agents to coordinate long-running jobs. If you are building multi-agent systems, A2A gives you a clear way to manage asynchronous work.
  • Read more

4. 7 AI Agent Skill Patterns Every Programmer Should Know — Medium

  • Why read: A guide to breaking large AI prompts into modular skills to get better results from coding agents.
  • Summary: Stuffing context windows with massive prompt rules usually leads to broken code. A better approach is using modular skills: short, single-purpose Markdown files that agents load only when needed. The author outlines seven skill patterns, such as tool wrappers, generator templates, and automated reviewers. They also cover "inversion" skills, which force the agent to ask clarifying questions instead of guessing, and "meta-skills" that manage loading the others. This setup lets engineering teams enforce code quality and architecture constraints without confusing the model.
  • Read more

5. A Primer on Compute Markets — X (formerly Twitter)

  • Why read: Explains how compute is turning into a traded commodity market with futures, spot pricing, and derivatives.
  • Summary: Data center spending is expected to pass $1 trillion in 2026, and compute is financializing just like oil and power did. Driven by demand for open-weight models and self-hosted AI, buyers want cheaper options than standard API tokens. The market is shifting from direct hyperscaler contracts to spot markets, loans backed by GPUs, and futures exchanges. Startups are building the plumbing for this, such as OTC desks and price indices. Anyone managing AI infrastructure costs needs to track how these markets develop.
  • Read more

6. [AINews] Claude Fable/Mythos 5.1: new SOTA model, 75% cache price cut but 70% more output tokens — Substack

  • Why read: Covers the performance, pricing, and initial user feedback on Anthropic’s Claude Fable 5.1 and Mythos 5.1 models.
  • Summary: Anthropic released Fable 5.1 and Mythos 5.1 for agentic coding and knowledge work. While cache read prices dropped 75%, benchmarks show the actual cost per task rose 20% because the models output more tokens. The models perform well on evaluations like Terminal-Bench-Science. User reactions are mixed: many like the improved planning and shorter answers, but note issues with rate limits and aggressive enterprise safety filters. There is evidence both models use the same weights but apply different safety thresholds.
  • Read more

7. Inside the multiplayer AI setups at Mintlify, LangChain, and Buffer — Substack

  • Why read: Looks at how marketing teams are building shared, multiplayer AI systems instead of working in isolated chat windows.
  • Summary: Marketing teams are shifting from solo ChatGPT sessions to shared AI setups. Buffer stores its strategic context in Notion and keeps its execution skills in a shared GitHub repo linked to a Claude plugin. This lets everyone on the team build and run AI workflows from the same knowledge base. They use MCPs to connect Claude to their internal tools, automating things like monthly reports, link checks, and partner applications. This shared approach lets teams scale their AI use without fragmenting their business context.
  • Read more

8. I accidentally made an agent factory — X (formerly Twitter)

  • Why read: A developer's account of accidentally building a multi-agent system that writes and tests its own software.
  • Summary: While testing Fable 5.1, the author spun up agents for different projects and gave them a Slack channel. When dozens of agents created chaos, the developer added a hierarchy: "lead" agents assigned tasks, and a "General Manager" handled conflicts and build priorities. They deployed specialized agents just for UI testing, which filled the backlog with bugs, and spun up an internal tools team to optimize token usage. The experiment shifted the developer's role from writing code to managing a team of AI workers, showing what it looks like when individuals run their own software factories.
  • Read more

9. AI Productivity Doesn't Mean What I Thought It Means — X (formerly Twitter)

  • Why read: Argues that AI doesn't necessarily save time, but instead raises the baseline quality of human work.
  • Summary: The author expected AI assistants to cut down the time it took to publish essays. Instead, the total amount of editing stayed the same. Looking back at five years of content, they found that working with AI improved the writing itself. Because the AI handles structural edits and catches bad drafts early, the writer can focus on rhythm and argument. In knowledge work, AI efficiency might not mean working fewer hours; it means getting better results for the same amount of effort.
  • Read more

10. What AI agents really think about your pricing — Substack

  • Why read: Shows why AI search engines misread B2B pricing pages and how to fix it.
  • Summary: As buyers use LLMs for software research, AI engines consistently get pricing wrong. An analysis of top cloud companies shows that AI rarely cites official pricing pages. Instead, it pulls from third-party sites or Reddit, which often have inaccurate or negative data. The problem is that many pricing pages hide details behind JavaScript, interactive sliders, or strict robots.txt files. To fix this, companies need to use server-side rendering and publish clear text documentation, FAQs, and edge cases. If AI crawlers can't read your pricing data, you lose control of what buyers see.
  • Read more

11. TBM 437: AI and the Recontextualization Tax — Substack

  • Why read: Looks at how AI can automate the busywork of translating ground-level data into executive reports.
  • Summary: Teams waste a lot of time on the "recontextualization tax": formatting raw, complex data to match what different managers and executives want to see. In the past, teams just ate the cost or avoided tools that couldn't handle multiple audiences. AI can bridge this gap by automatically converting raw workflow data into whatever format a stakeholder prefers. Teams can stick to their own messy, functional workflows, while the AI generates clean roadmap views for leadership. Offloading this translation work saves time and stops organizations from flattening reality just to make it fit on a slide.
  • Read more

12. Zero Data Retention — X (formerly Twitter)

  • Why read: Explains why enterprise customers are demanding Zero Data Retention (ZDR) from AI providers.
  • Summary: Enterprise buyers are shifting traffic away from models like Anthropic's Fable 5 because of its 30-day data retention policy. They prefer Zero Data Retention (ZDR), where prompts and outputs are deleted immediately, removing the risk of leaks, internal snooping, or subpoenas. Large customers care more about data privacy than helping providers train future models, giving CISOs veto power in AI procurement. AI builders need to prioritize confidential inference and guarantee ZDR in their vendor contracts. Forced data retention is becoming a dealbreaker for enterprise sales.
  • Read more

13. We Don’t Understand What Emerges Along the Intelligence Curve — X (formerly Twitter)

  • Why read: Argues that future leaps in AI capability will happen across networks of agents, not just inside individual models.
  • Summary: Scaling laws suggest AI gets smarter in a straight line, but new capabilities like tool use and in-context learning actually appear in sudden jumps. The author suggests the next jumps won't happen inside isolated models. They will emerge in multi-agent systems where specialized models share tools and divide labor. Just as humans advanced through markets and the printing press, networked AI agents could learn continuously through shared memory. Focusing only on benchmarking single models misses the point. The next level of intelligence will probably look like a complex organization, not a giant standalone brain.
  • Read more

14. Keeping Open Source Open in the Age of Coding Agents — X (formerly Twitter)

  • Why read: Explores how AI coding agents are stripping away open-source license attributions and what it means for developers.
  • Summary: Coding agents routinely read open-source code and rewrite it into commercial projects without keeping the original MIT licenses or attributions. This forces maintainers to hunt down license violations while companies use their work for free. To keep open source viable, engineering teams need to require AI agents to track sources, keep copyright notices, and flag borderline cases for human review. If agents keep erasing software history, researchers will start gating their work. We need to train agents to handle attribution as carefully as human engineers do.
  • Read more

15. What It Took to Build a Self-Learning Content Engine — X (formerly Twitter)

  • Why read: A case study of how Atlan replaced their manual content marketing with an AI agent network.
  • Summary: As LLMs disrupted organic search, Atlan rebuilt their content strategy. They moved from manual writing to an AI pipeline, and finally to Helix, a fully autonomous agent. Helix runs on a shared system that pulls data from customer calls and social trends to plan, draft, and edit posts. It uses Claude for reasoning, Gemini for search, and Grok for fact-checking. This setup cut their content budget and sped up publishing, freeing the marketing team to focus on strategy instead of drafting posts.
  • Read more

Themes from yesterday

  • Multi-agent orchestration: Workflows are moving from single-model pipelines to orchestrator-worker networks. This unlocks new system-level capabilities without relying entirely on larger models.
  • The financialization of compute: Compute is becoming a globally traded commodity. We're seeing spot markets, futures, and GPU-backed loans to meet the demand for self-hosted AI.
  • Zero Data Retention: Data retention is a dealbreaker for enterprise adoption. Large customers are routing traffic away from providers who log data and toward those offering immediate deletion.
  • Context over prompt engineering: Teams are abandoning massive prompts. Instead, they use shared repos, message histories, and modular skill files to give agents reliable context.
  • The open-source provenance crisis: AI agents are rewriting open-source code into commercial apps without keeping license attributions, pushing maintainers to reconsider how they share their work.