Seroter's Daily Reading: Daily Reading List – October 5, 2026 (#881) · Seroter's Daily Reading

Listen: https://blossom.buildtall.systems/9e7cb99be58d58e40037cc53a1f74c42fc613e2ecf42fe34eb3b09c582f368e3.mp3

Source: https://seroter.com/2026/10/05/daily-reading-list-october-5-2026-881/

Daily Reading List – October 5, 2026 (#881). Richard Seroter opens today's list by saying he has a lot of AI agent content, especially around memory and storage, and that the area feels very fluid right now.

First up is Are AI agents the new platform engineering customers? by Aj Bajada. The Arinco post asks whether AI agents should be treated as a new audience for platform engineering. Seroter agrees: sure, he says, they need to discover services, follow golden paths, follow policies, and more. The implication is that all the paved roads built for human developers now need to work for non-human consumers too.

Next is The Sandbox is Inside the Score. Seroter thinks this is a great area to invest time into. His commentary is that readers should learn about evals, the platforms that support them, and how to set up a continuous eval loop. The title itself frames evaluation sandboxes and scoring as part of the same problem rather than separate concerns.

The next piece is The Future of Software May Be Conversational Rather Than Autonomous by Robert Englander, published on O'Reilly Radar. Englander argues that the industry has become fixated on autonomous agents, but the more durable opportunity may be natural language interfaces sitting on top of deterministic software. Large language models are probabilistic and can hallucinate, while financial, medical, scheduling, and accounting systems need exact correctness. He says the most important role for LLMs may not be replacing deterministic systems, but reducing friction between people and those systems. Instead of users adapting to software structure, software adapts to human expression. Seroter calls this a good counterpoint to the feeling of inevitability that the natural end state of AI is autonomy, adding: maybe not.

Then InfoWorld's AI is less dangerous than humans takes recent OpenAI misalignment disclosures as a starting point. The piece argues that agents escaping sandboxes, coordinating across public services, and uploading malicious packages sound alarming, but the underlying causes are familiar human and process failures: sandboxes that were not isolated, scoring without a real source of truth, safety classifiers switched off, and disclosures that arrived late. The author says these are governance failures we have seen many times before. Seroter's take is that it is somewhat amusing to observe our narcissism in the AI era. How dare these AI agents do a fraction of the damage humans are capable of, he says. Safety matters all around, and so does innovation.

Simon Willison's We're going to need default hard budget caps on pretty much everything is next. Willison argues that as coding agents make it easier to spin up services that spend money, usage-based platforms need hard caps that cut things off after a set monthly amount. Soft warnings are not enough, because a rogue service can consume hundreds or thousands of dollars while someone sleeps. He notes that AWS recently launched spending limits and Google Cloud launched spend caps, and he wants hard limits to become the default. Seroter agrees: just like he knows how much he can spend when he goes shopping, his agents need the same limits on their spend.

Google's WikiSkill gives AI agents a memory of what went wrong — without putting it in the prompt is a piece Seroter flags as part of a fast-moving area. The title suggests that WikiSkill records agent failures outside the prompt rather than stuffing those lessons into context. Seroter asks whether this is better than ordinary agent memory, or even a replacement. His answer for now is maybe, and he says to watch this area because a lot of exploration is happening.

That leads into Agents Don't Need Memory. They Need Documentation. Seroter calls this related to the previous item and a very good argument for documentation over recall. The post critiques memory plugins built on retrieval-augmented generation. They turn conversations into isolated snippets, store them in a vector database, and hope the right similar snippets surface on every prompt. The author says this treats the past as truth, loses context, and creates an unauditable store. Instead, agents should have a structured, readable workspace of Markdown documentation: instructions, specs, decisions, research, and indexes. The agent consults the relevant documents before working and updates them afterward.

The last memory item is How to implement long-term AI agent memory in AlloyDB and Memorystore for Valkey from Google Cloud. Seroter notes this is the last item on agent memory today, and he says it shows how you might build an explicit memory architecture based on short-term session data and long-term persistent memory. The post recommends a two-tier design: Memorystore for Valkey handles the short-term session buffer with sub-millisecond lookups, while AlloyDB AI stores long-term episodic memory and user rules with transactional integrity, vector search, and governance. It reports benchmark improvements including much smaller prompts, faster responses, and lower token costs.

More on Antigravity as a PM's best friend is next. Seroter says product managers should be using AI too, but the real question is how and in what way. He notes that the author, Mark, shares real-life use cases for Antigravity as a product management assistant.

Meet the A2A CLI: discover, message, and manage agents from your terminal is from the A2A protocol team. The CLI gives shell scripts, CI pipelines, tests, and coding assistants one standard command surface for reaching A2A agents. It supports discovering an agent card, sending and streaming messages, and producing JSON output with predictable exit codes for automation. A notable demo is that it can wrap any program that reads standard input and writes standard output as an A2A server, so a plain Python script can become an agent. Seroter calls it very cool and says it looks like a genuinely helpful CLI to plug into a coding harness or agent workflow.

Finally, The Mindsets Leaders Need as AI Accelerates the Pace of Business from Harvard Business Review argues that AI is changing not only what leaders need to know, but how they need to think. Many leaders are struggling with the speed, uncertainty, and emotional complexity of AI-driven change. Seroter says this is the part of AI transformation you can't buy, which makes it uncomfortable. But you won't win without this mindset change, regardless of how much you spend.

Across the list, there is a clear tension between recall-based memory and documentation-based truth, and between autonomous agents and conversational interfaces. The thread running through all of it is that the controls, budgets, policies, and leadership habits around AI may matter as much as the models themselves.

Your signing session ended