Architecture AI Updates: November 4, 2026

1. Netflix’s Performance Agent Profiles Production, Proves Fixes on Canaries and Shares What It Learns Through a Markdown Pattern Catalog

ByteByteGo. In a November 3 post based on Netflix’s public talks, including a P99 Conference presentation, ByteByteGo described the autonomous performance agent system that Netflix built because developers using AI ship code faster than its small group of performance experts can profile it. The pipeline gives profiler output (async-profiler or Java Flight Recorder on the JVM, py-spy for Python, pprof for Go) to an LLM and asks it to match known anti-patterns against real call stacks: quadratic loops, work repeated inside loops that could run once, allocation on hot paths, N+1 queries. Strict package-prefix rules stop the model from spending reasoning tokens on garbage collection, kernel syscalls and third-party framework overhead that the team cannot change. In production services the agent found four classes of problems: an immutable map rebuilt inside a Java Streams reduction that cost 8.8% of CPU; a regex substitution run on every request that took 46% of a service’s CPU, which a lookup table replaced at 22 times the speed; a HashSet created on every call that accounted for 47% of heap collections; and a metric counter looked up from the registry on every call, which took 66% of CPU against 0.3% for the business logic and appeared in three services.

Each fix must pass the existing unit, integration and contract tests. It then runs as a canary beside an identical baseline, and CPU, median and P99 latency and allocation rates are compared without any model involved. Only after that does the system open a pull request containing the diff, the original flame graph and the canary report, and a person decides whether to deploy it. To catch problems before production and stop agents finding the same bug again, findings go into a Pattern Catalog: a Git repository of plain Markdown files, with no vector database, that people review like any other code. The catalog is injected into system prompts. In CI, a regex or symbol match flags new pull requests that resemble known anti-patterns, and an LLM then comments inline on the diff with the impact and a suggested fix. Netflix places itself at level 2 (orchestration) of a four-level maturity scale. Its long-term goal is level 3: fully autonomous agents that explore and benchmark alternative implementations on their own. Source

2. Thoughtworks’ Nimisha Asthagiri Proposes a “Bounded Canvas” Instead of a Harness for Coding Agents

InfoQ published a QCon AI talk on November 3 by Thoughtworks technical advisor Nimisha Asthagiri, formerly chief architect at edX. She argues that “harness engineering”, the stage after prompt and context engineering in which developers keep improving the guides and sensors around a coding agent, treats agents too narrowly as something to restrain from outside. She proposes a “bounded canvas” instead. On the feedforward side, the canvas sets boundaries, scaffolding and starter kits, chosen constraints and lessons from earlier prototypes. On the feedback side it has four layers: structural (type systems and sandboxes), reactive (backpressure, rate limits and linters), dynamic (guardrails that detect architecture drift, drift from the spec and budget overruns) and reflective (periodic reviews and retrospectives).

She treats cognitive debt as unavoidable once people stop writing the code themselves. Agents should therefore leave a trail, which she calls “stigmergy”: architecture, product and organization decision records, plus guidance on which decisions need human review depending on how reversible they are. For a life sciences client she reused a product engineering operating model (define, build, review, with mission, goals, KPIs and key results passed down as context). The agent wrote these as Markdown files, built a dashboard showing decay, staleness and drift, and was put through deliberate disruption in the spirit of chaos engineering, such as removing code or a tool it expected to have. In Q&A she recommended keeping only a few maps up to date, such as C4 levels 0 and 1 and a folder-structure map, rather than long CLAUDE.md files that enter the context on every turn. She also said her own strategy documents over-constrained the agent at first. Source