Listen: https://serve.buildtall.systems/7e4a1d1f9f865b60b477266272855fca2c3ec221c47c8a88909576ffa69fe950.mp3
Source: https://seroter.com/2026/11/12/daily-reading-list-november-12-2026-908/
This is Seroter's Daily Reading, covering Daily Reading List – November 12, 2026 (#908). Richard Seroter opens by saying it's another great day of learning and building in Tokyo, and that there's an insane number of things happening in tech this week. The list moves across decision APIs, engineering leadership, agent memory, enterprise surveys, certifications, and the changing development workflow.
First up is Gemini Decisions API: Fast, calibrated multimodal decisions, from Google. This introduces the Gemini Decision 1 model and a new endpoint for fast, closed-set decisions. Instead of generating answer text token by token, the model scores candidate options in a single forward pass, returning probabilities and confidence scores. It handles up to a thousand independent yes/no, multiple-choice, or rubric-scoring questions in one request, across text, images, audio, video, and PDFs. Google reports end-to-end latency of 150 to 250 milliseconds and pricing of 10 cents per million input tokens, with zero output tokens billed. The post also describes internal uses like dynamic model routing, rapid design iteration in Stitch, emoji selection, and a safety layer for MCP tools. Seroter's take is direct: this is a decision model, and you can send in all sorts of input along with independent questions and get back probabilities and confidence scores.
Next is Beyond Writing Code: Taste, Skills, and Tech Leads. Seroter's entry frames the piece this way: we're all tech leads now, but where does the human work come in? He adds that Daniela has a good take. The title signals a focus on the non-coding side of technical leadership: taste, skills, and the role of tech leads.
Then comes The future of Software Engineering is boring, by Alex Pliutau. Pliutau argues that AI coding has changed the profession from one where programmers carefully crafted and understood code, to one where generated code piles up faster than it can be reviewed. He describes three problems: how do we learn when the path from problem to solution skips the brain; codebase erosion as AI-generated code bloats systems and review is skipped; and noise, management pressure, and slop. The result, he says, is that software engineers have become caretakers of code that nobody really understands. Seroter says he gets what Alex is saying, but he is enjoying software engineering now more than ever. At the same time, Seroter warns that we will need to actively adjust to keep learning and not become meat proxies.
The next article is The components of agent memory, published by Anthropic. It examines how AI agents can maintain knowledge between sessions. The central idea is to use offline compute, sometimes called dreaming, to reflect on session transcripts and consolidate experiences into structured notes, while catching errors. The post compares this to human memory consolidation during sleep. It walks through experiments with Claude models, including a benchmark where models wrote memory notes and a Pokémon Red task where better models wrote fewer, more strategic memories. Dreaming corrected a mistaken location note and allowed the agent to progress further. The authors also found that steering older models to use memory helped, while stronger models benefited more from freedom to judge what was salient. Seroter's commentary picks out the core: AI agents can consolidate experiences into structured notes and catch errors by dreaming on session transcripts.
We gave agents better instructions, from Mintlify, describes a practical fix for agent hallucinations. The team built a CLI command called mint instructions that prints current, task-specific guidance. In the first benchmark, Claude Sonnet passed only 16 percent of runs because agents never ran the command. After adding one pointer in the repository's root AGENTS.md file, agents ran it before their first edit in 30 of 30 tests, and Sonnet's pass rate reached 100 percent. Task-specific CLI guidance beat pointing agents to documentation on every model tested. The article also reports false claims about Mintlify dropped from 0.69 per run to 0.02. Seroter calls it a cool idea from the team at Mintlify, saying this new command gives agents guidance that greatly reduces hallucinations.
New survey finds enterprises trapped between legacy software and AI modernization reports on a Google Cloud and Hypothesis Group survey of 1,250 enterprise IT and business decision makers. The survey describes legacy on-premises software as an innovation tax. Nearly half of respondents say legacy software has delayed, reduced, paused, or canceled AI and technology initiatives. Patch application and maintenance create a zero-sum game between keeping old systems running and modernizing with AI. The report also highlights vendor lock-in and enterprise monoculture as sources of operational risk, and argues for genuine choice and interoperability. Seroter's view is that companies are held back by previous investments, which is nothing new, but he says the urgency is higher now and calls the data interesting.
Anthropic and OpenAI are coming for the development workflow, from InfoWorld, looks at how frontier labs are extending beyond models into the entire developer toolchain. The article cites OpenAI's agreement to buy Pierre, the company behind Code Storage, and Anthropic's redesigned Projects as examples. It points out that labs face pressure to maximize profitability and own more of the developer workflow. The author argues that developers should be wary of turning a preference for this quarter's best model into a long-term commitment to one supplier. The piece calls for neutral spaces like GitHub where competing agents can work together, and a practical way to change which agents a team uses. Seroter says he hears Matt's point, and adds that you probably don't want to pick fundamental infrastructure such as source code storage based on which model you like this month.
How Salesforce Scaled Engineering AI Adoption Without Stifling Innovation is an experience report from Salesforce engineering. The page's available text highlights a related piece, Using AI to Build AI: How Agentforce Engineers Scale Automation Without Sacrificing Reliability, by Fereydoon Vafaei. Seroter says he likes these types of experience reports coming from big shops, not just small startups with no preexisting tech commitments. The framing is that large organizations have existing systems and commitments, so their adoption stories are particularly useful.
Don’t just prompt AI — architect it: Introducing the Professional Agentic Architect Certification, from Google Cloud, announces a new certification for building production-ready agents. It uses a hybrid format: a proctored multiple-choice exam on system design, cost optimization, security, and governance, plus hands-on labs where candidates build, test, and debug live agentic workflows. The exam covers five domains: building agents with low-code tools, using coding agents, developing custom agents on the Gemini platform, evaluating and deploying agentic workflows, and securing and governing them. Google says the certification grew out of internal training for forward deployed engineers and partners. Seroter says the certification is a legit test of knowledge, and points readers to the topics covered as evidence.
AWS gives developers a free local stand-in for its cloud, built from scratch by AI agents is a GeekWire article. Seroter's reaction is that an agent-built emulator for AWS is pretty cool work, and he points to documentation for it. The title indicates AWS has a free local emulator built by AI agents.
Seroter's linked documentation page, titled Some docs here, is the getting started guide for AWS Emulator. It describes four ways to start: with an AI coding agent using the aws-emulator skill, with a Linux package, with Docker, or inside a Rust test process in embedded mode. The Linux and Docker paths start a local process listening for HTTPS requests on port 9297 on localhost, and you point the AWS CLI or an AWS SDK at that endpoint with environment variables. The page also directs readers to setup prerequisites before beginning.
Finally, Redesigning the Dashboard: Cultivating Trust Over Gamification in AI Adoption, from IEEE, is the last item. Seroter says the Google developer experience team learned some things about token tracking dashboards. The title points to designing dashboards that cultivate trust rather than gamify AI adoption.
Across the whole list, the recurring tension is between speed and control: decision APIs make agent choices fast, Mintlify shows how better instructions reduce hallucinations, Anthropic explores memory and dreaming, and the development workflow pieces warn against locking into a single model vendor. Seroter seems excited but also mindful that the hard work is human: taste, learning, trust, governance, and choosing infrastructure for the long term.
- Daily Reading List – November 12, 2026 (#908)
- Gemini Decisions API: Fast, calibrated multimodal decisions
- Beyond Writing Code: Taste, Skills, and Tech Leads
- The future of Software Engineering is boring
- The components of agent memory
- We gave agents better instructions
- New survey finds enterprises trapped between legacy software and AI modernization
- Anthropic and OpenAI are coming for the development workflow
- How Salesforce Scaled Engineering AI Adoption Without Stifling Innovation
- Don’t just prompt AI — architect it: Introducing the Professional Agentic Architect Certification
- AWS gives developers a free local stand-in for its cloud, built from scratch by AI agents
- Some docs here
- Redesigning the Dashboard: Cultivating Trust Over Gamification in AI Adoption

