Seroter's Daily Reading — #767 (April 20, 2026) — Seroter's Daily Reading

Listen: https://blossom.nostr.xyz/e0728d9773db36c29c233994e04ed4f5483457b71f56c5e8b1d2f8e951ac1d98.mpga

Seroter's Daily Reading, Episode 767. April 20, 2026.

I'm in Las Vegas this week for Google Cloud Next '26, and spent the last two full days in various keynote rehearsals. Fun stuff with people I enjoy hanging out with. On with the links.

Leading off this week — a piece from Addy Osmani on The Agent Stack Bet. This one caught my eye because it cuts through a lot of the hype and gets specific about the problems that actually burn on-call rotations when you ship agents to production. Addy calls out what he calls excessive agency — autonomous systems given broad permissions to get things done, then left to discover at runtime that a schema drifted, an API changed, or a downstream service started returning data it wasn't supposed to. The agents mark tasks complete while leaving a trail of corrupted state behind them. The humans find out on Monday.

He lays out four architectural bets that every serious team needs to make in the next twelve months. First: agents need identities, not shared credentials. Most agents today are ghosts — they borrow a service account, inherit a human's OAuth token, and promise in a prompt to stay inside the lines. In a real enterprise environment, a promise in a prompt is not a policy. The fix is moving agent identity from the application layer down into the platform layer, so the agent has an unforgeable identity recognized at the network level and policy is enforced at the source. If the agent reaches for a database it isn't cleared for, the connection never opens.

Second: agents need universal context, not scraped windows. Right now teams are burning huge engineering hours on custom serialization, bespoke session stores, hand-rolled memory layers — just to keep an agent from forgetting its mission halfway through a multi-step task. And the context agents can get their hands on is usually siloed. A browser-based agent can see the open tab. A desktop wrapper can see the files a user dragged in. Neither can easily reason across the systems where the business actually lives — the CRM, the ERP, the data warehouse, the ticketing system — at the same time. Without fixing this, Addy argues the ceiling of agentic AI is just slightly better spreadsheet autocomplete.

Third: agents need to survive your laptop closing. The frontier has genuinely moved in the last six months — persistent task state, scheduled execution, multi-agent coordination are no longer aspirational. The question now is whether an agent can run for a week instead of an hour. Whether it can cross three handoffs, two credential rotations, and an approval gate without a human babysitting the session. Enterprise-grade autonomy requires durable, cloud-native execution with state and checkpointing that survives restarts, disconnects, redeploys, and model version changes by default. Anything less is a demo that happens to run for a long time.

Fourth: agents need platforms. The pattern he sees most often in strong teams is the saddest one — brilliant engineers draining their bandwidth into stack problems that don't differentiate their product. Custom memory, bespoke eval harnesses, homegrown observability, handwritten retry logic. The real value lives in domain reasoning and business logic. Everything underneath should be the platform you build on, not the plumbing you build. His bet is on open orchestration frameworks that let teams prototype on their laptop with the same building blocks they'll run in production, and cross that boundary without a rewrite.

Next up — a piece from A Smart Bear on How to Hire People Who Are Better Than You. If you don't hire people better than you, the organization gets bigger, not better. But how do you hire for something you don't understand? Jason Cohen walks through this with his usual clarity. You leave the interview wanting to implement their ideas — that's the signal. Genuine expertise produces specific, actionable ideas that feel tailor-made to your context. People who just talk a good game will say things that sound impressive but will leave you thinking is that really what grown-up marketing is, not hell yes that's what I want this company to be like. If they change how you think about your own company in an hour, imagine what they could do in a year.

He also suggests asking reference checks about culture fit rather than expertise. Specifically, ask: what is the ideal scenario under which this person thrives? And then ask the opposite: construct the scenario where they die inside, mess everything up, and piss everyone off. One more question he likes: what are their strengths that they're unaware of because they come so naturally to them — where they don't realize other people aren't like that too? These are superpowers. You're not looking for someone without weaknesses. You're looking for someone with a few special strengths that your company needs right now.

The focus throughout is on hiring people who elevate the whole organization, not just their silo. Great leaders are better at people issues, communication, decision-making, goal-setting, annual planning, org structure — things that are useful everywhere in the company. Those clues are available to you during the interview. The culture fit was available in the reference checks. It's not perfect, but that's what you do.

Moving to Google Cloud now. They've announced a unified Spanner Graph and BigQuery Graph solution. Spanner Graph handles operational workloads — bringing together graph, relational, search, and generative AI capabilities into a single database, backed by Spanner's unlimited scalability and strong consistency. You get integrated table-to-graph mapping so you can view and query operational data as a graph without duplicating it, and you can mix ISO-standard Graph Query Language with SQL in a single query to traverse both graph and tabular data together.

BigQuery Graph handles the analytical side — mapping existing BigQuery tables to a graph schema instantly, uncovering hidden relationships in your data warehouse without building ETL pipelines or moving a single byte. Customers like WPP are using this for identity resolution across millions of entities, and Yahoo is using it to power real-time intelligent decision-making across their advertising platform. The unified approach eliminates data silos and accelerates time-to-insight.

From Harvard Business Review, a piece called When Your Ambition Starts to Exhaust You. Earlier in your career you volunteered for hard projects, stayed late, and prided yourself on delivering. It worked. You rose through the ranks, built a reputation for getting things done and became a leader. That was then. Now, the ambition that once energized you exhausts you. And the standards you set feel like a treadmill you can't step off. I'll leave the full read to those who need it, but the piece offers five questions to ask yourself when you hit that wall.

Over to Google Cloud documentation — a guide on Multi-Agent Private Networking Patterns. This one is very good for anyone working through multi-agent setups. The architecture uses a VPC network with private IP addresses, and agents communicate using the Agent2Agent protocol and the Model Context Protocol. The design supports agents hosted in Vertex AI Agent Engine, Cloud Run, GKE, on-premises data centers, or in other clouds, with Private Service Connect endpoints and Cloud Run Direct VPC egress to route traffic securely. Cloud NGFW governs traffic through the VPC, and additional layers provide controlled internet egress through Secure Web Proxy and API service access policies through VPC Service Controls perimeters.

The root agent acts as top-level workflow orchestrator, delegating tasks to specialized subagents. Each hosting platform has different private connectivity mechanisms: Vertex AI Agent Engine uses Private Service Connect interfaces, Cloud Run uses Direct VPC egress, and GKE clusters reside directly in the VPC with local subnet IP addresses. Gemini Enterprise apps are managed resources outside the VPC but within Google's network, with the built-in chat interface delegating to registered agents through the discovery service.

Finally, from TechCrunch — a piece on Tokenmaxxing and whether it's making developers less productive than they think. There's an old saw in management: what you measure matters. And typically you get more of whatever you're measuring. Enormous token budgets have become a badge of honor among Silicon Valley developers, but measuring an input to the process makes little sense when you presumably care more about the output.

The data from companies in the developer productivity insight space tells a consistent story. Engineers are seeing code acceptance rates of 80% to 90%, but they're missing the churn that happens when engineers revise that code in the following weeks, which drives the real-world acceptance rate down between 10% and 30% of generated code. GitClear found that regular AI users averaged 9.4x higher code churn than their non-AI counterparts — more than double the productivity gains the tools provided. Faros AI found that code churn increased 861% under high AI adoption. Jellyfish collected data on 7,548 engineers and found that the engineers with the largest token budgets produced the most pull requests, but the productivity improvement didn't scale. They achieved two times the throughput at 10 times the cost of tokens.

In other words, the tools are generating volume, not value. One common finding is the difference between senior and junior engineers, with the latter accepting far more AI-generated code and dealing with a larger amount of rewriting as a consequence. Even as developers work to understand exactly what their agents are up to, they don't anticipate turning back anytime soon. This is a new era of software development, and you have to adapt. It's not like it will be a cycle that will pass.

Wrapping up this week's batch — a couple of threads running through the reading. One is infrastructure: whether that's the agent stack, multi-agent networking patterns in Google Cloud, or the unified Spanner and BigQuery Graph solution, the story is about getting the foundational plumbing right so you can stop fighting it and get back to the actual product. The other thread is productivity measurement — whether that's hiring people who are better than you, understanding what token budgets actually cost, or managing ambition before it exhausts you. The common thread is that inputs and outputs often look very different, and measuring the wrong thing leads you astray no matter what you're building.

That's episode 767. Links and show notes are live at seroter.com.

The Agent Stack Bet — Addy Osmani on four architectural bets for production agents: identity, universal context, survivable sessions, and platforms.
How to Hire People Who Are Better Than You — Jason Cohen on evaluating candidates who know more than you, including reference check questions that reveal culture fit and superpowers.
Unified Spanner Graph and BigQuery Graph Solution — Google Cloud's combined operational and analytical graph database offering.
When Your Ambition Starts to Exhaust You — HBR on recognizing and responding to ambition-driven exhaustion.
Multi-Agent Private Networking Patterns in Google Cloud — Architecture guide for VPC-based multi-agent deployments using A2A and MCP protocols.
Tokenmaxxing Is Making Developers Less Productive Than They Think — TechCrunch on token budgets, code churn data, and the gap between AI tool adoption and actual productivity gains.
Designing Synthetic Datasets for the Real World — Google Research on mechanism design approaches to synthetic data generation.