Weekly Review — Nov 2–8, 2026

On any single day the common thread is easy to miss. An average, a total or a two-column table can look fine while the part you care about gets worse, so most of this week went into finding the number underneath it.

This Week on the Blog

Which Agent Platform? Bedrock Managed Agents, AgentCore Harness, or Runtime

Bedrock Managed Agents, powered by OpenAI, went into preview on September 29, and it is tempting to read it as Managed Agents versus AgentCore. The docs describe three options, and what separates them is who runs the agent loop: OpenAI’s Codex harness operated by AWS, the AgentCore harness built on Strands Agents, or your own code on AgentCore Runtime. I gave the same small shell task to Managed Agents and to the harness and compared where the command ran, what got created by default and which OpenAI models each Region offered. A side-by-side table, three example workloads and a checklist close it out.

Why Does a Curveball Curve?

Two popular answers say the air grabs the ball or that the curve is an optical illusion, and neither holds up. The spin drags a thin layer of air around the ball, so the flow is faster and the pressure lower on one side, and the ball gets pushed toward that side all the way to the plate. Flip the spin and the curve flips with it, which you can try in the interactive demo. Episode eight of Lunch Break Physics also covers free kicks, golf backspin, rotor ships and why the “late break” really is partly in the batter’s eyes.

Lambda Web Functions: What Changes When Waiting Stops Costing CPU

Lambda Web Functions, launched on 12 October, run a standard Node.js server with no event-to-HTTP adapter, and one environment serves up to 128 requests at once. You pay for requests, the CPU your code actively uses and the memory each environment holds. In AWS’s own pricing example a 3-second streaming request is billed for 200 ms of CPU, and a small test pointed the same way: 1,584 waiting requests used 0.27 seconds of CPU. CPU-bound routes, very short high-volume requests and quiet traffic gain much less, and every endpoint is reachable from the internet, so the post walks through CloudFront with origin access control.

AI Usage Metrics Make a Good Smoke Detector and a Bad Scoreboard

I once projected a token leaderboard at a customer hackathon, knowing tokens are a weak metric, because shy participants needed a nudge. For three days that worked fine. Turn the same chart into a standing metric with careers attached, though, and people start optimising the number. Amazon’s engineering memo suggests watching the 10th and 90th percentile of adoption and asking why, which I read as a smoke detector. The post adds four rules (teams not individuals, an outcome next to usage, no target, nothing in reviews) and a five-part write-up for teams where AI doesn’t help yet.

Golden Questions: Every Agent Change Is a Release

A new model version, a prompt edit or a changed tool contract each changes the agent, so each deserves a release gate. The gate is a fixed set of golden questions with known-good answers, run several times per change, with three checks: an overall floor, a limit on how far any category may drop, and zero safety failures. In a simulation two candidates beat the baseline average and were both blocked, one for a collapsed category and one for a leaked email address. Run against three Nova models on Bedrock, the same suite blocked the release with the highest average and turned up a bad golden question along the way.

The Thread This Week

The golden questions post makes the point most directly: a model upgrade raised the average and broke multi-step questions, and only a per-category limit caught it. Thursday’s post makes the same move with AI adoption, looking at the tails of the distribution instead of the mean. On Wednesday the question is which meter runs, wall-clock time or active CPU, and in AWS’s streaming example the request is open fifteen times longer than its CPU meter runs. And Monday’s comparison only becomes useful once a third column appears, because “Managed Agents or AgentCore” hides who owns the loop.

Further Reading

Things worth your time that did not get their own post, all public:

• Agentic Software (Zhenfeng Cao, arXiv): Zhenfeng Cao’s paper went up as “The End of Software Engineering” and was retitled six days later, now describing agentic engineering as an expansion of the discipline. The more useful evidence sits in a benchmark it cites, SWE-Milestone: across 12 frontier models and 4 agent frameworks, overall performance scores fall from above 80% on isolated tasks to about 38% when the agents have to keep evolving the same codebase.
• AI-DLC: The AI-Driven Development Lifecycle (AWS Humans in the Loop): Shaown Nandi and Anupam Mishra argue that coding is only about 30% of a developer’s time, so doubling coding speed shortens end-to-end delivery by only about 15% if everything else stays the same. AI-DLC goes after what they call the coordination, testing and translation tax, and measures speed and quality together.
• How to Deslopify Your AI (The AI Daily Brief): Defines slop as work where the person who should have done the thinking skipped it and moved the cost to whoever receives it. The practical advice: tell the tool your position and how sure you are, ask for several different directions, and decide in advance where you will step in, because a polished draft is the hardest one to throw away.
• Are Vector Databases Still Needed? (IBM Technology): A four-level pyramid for retrieval: long context, keyword search run by an agent in its loop, a vector index in the database you already have, and a dedicated vector database. Climb only when the semantic gap or the scale demands it, and remember that switching to a different embedding model generally means re-embedding the indexed corpus.
• The 2026 Software Factory Infra Stack (Latent Space): Zach Lloyd describes a software factory as an agentic system defined in code per product surface, with every run stored and scored. By Warp’s own numbers, after tuning the factory its cost per pull request fell from about $80 to $15–20, while throughput rose 40% in a month with the revert rate unchanged.
• How Trajectory Is Building Products for Continual Learning (Sequoia Capital): The loop runs from production traces to a replayable environment, a reward, training and redeployment. The scarce input is “what good looks like”. Trajectory’s approach is to label about 100 representative traces out of 10,000 with a rubric and train a grader that scores the rest.
• Building Legal Agents When You Can’t See the Data (Latent Space): Legora’s CTO on growing from a legal research tool into an agent that drafts, edits and orchestrates work, with forward deployed engineers who, as he puts it, build integrations for law firms without touching client data. He also says most legal work doesn’t need the most capable model.

Until Next Sunday

If you take one thing from this week, pick the average you trust most for your agent or your team and split it by category. I’d look at whichever category dropped first.

This is the Weekly Review, and it also goes out as Sunday’s newsletter.