Architecture AI Updates: November 7, 2026

1. Nicole Koenigstein Picks Coding Harnesses by the Layer They Control, and an Isolated Codex Verifier Caught a Bug Claude Code Contested

Nicole Koenigstein, writing on O’Reilly Radar on November 6, argued that coding harnesses should be chosen by which layer of control they hand over rather than by who built them. In her map, Claude Code gives control inside the run (CLAUDE.md instructions, skills, subagents and hooks that can explain a refusal or return a compiler or policy failure to the agent), Codex gives control around execution (noninteractive codex exec runs under an OS-enforced sandbox with restricted writable paths and schema-constrained output), Pi gives control over the loop itself (tools, context, state and hand-offs can be replaced) and DeepSeek’s harness gives control through composable plug-ins; whatever a harness does not provide, such as verification, durable state, repair routing or an enforced execution boundary, stays with the team. For a Python 2.7 to Python 3 migration she composed three of them: Claude Code built the port with implementation and review subagents, isolated Codex runs on clean checkouts verified it without the builder’s instructions, Pi ran the outer lifecycle with deterministic rules for bounded repairs, replanning and human escalation, and frozen acceptance checks (Playwright plus Python 2/3 differential tests) ran outside the agents so neither builder nor verifier could redefine success.

On a test application seeded with similar traps, both the composed setup and a single Claude Code session passed all 25 hidden behavior tests, but the single session silently merged the dictionary keys u’é’ and ‘é’ into one Python 3 string, losing an entry once the data was saved, a defect also missed by Claude’s own tests, hundreds of recorded behavior comparisons and her own hand-written reference port. Two isolated Codex verifiers reproduced it, Claude contested the finding, and a fresh Codex run given only the finding and Claude’s argument upheld it before Pi routed it into a bounded repair. Her conclusion is that independent verification helps because the second context is not anchored to the builder’s implementation history, not because the second model is automatically right. Source

2. Markus Eisele Lists What a Coding Assistant Must Show Before Developers Can Trust Automatic Model Routing

Markus Eisele, writing on O’Reilly Radar on November 5, argued that automatic model routing is worth accepting when it buys more capable assistance from a fixed budget, but that comparing routing policies means following a task through its retries, corrections and developer time rather than per-request cost. He cites RouteLLM, whose preference-trained routers cut cost more than twofold with minimal quality loss on its benchmarks, and the SWE-Router approach, in which a cheaper model makes a short first attempt and the router uses what happened to decide whether to escalate, as a better fit for coding agents because a task becomes clearer as the agent reads code. Using a Java order service that creates duplicate orders after a timeout and retry, he shows that a stronger model only helps if the harness gives it access to the receiving service’s code, and that a proposed fix that never checks whether the first order was saved may reflect missing information rather than a weak model. His minimum for a routed assistant: let the developer ask for more budget or a stronger model on a hard investigation, show which files the agent examined and which claims were checked versus assumed, keep unresolved questions visible, state which code version tests ran against, and show token usage next to the actions that consumed it. For teams picking a policy, he recommends trials on everyday tasks with similar starting conditions that count corrections and developer time, and treating repeated retrieval of the same information as a sign the harness needs better context handling. Source

3. Yelp Renders Its AI Assistant’s Conversation From the Server but Keeps Navigation and Input Native

Yelp uses a hybrid interface architecture for Yelp Assistant, InfoQ’s Leela Kumili reported on November 6: navigation and message input stay native on iOS and Android, while the conversation area is rendered by CHAOS (Content Hosting Architecture with Optimization Strategies), Yelp’s server-driven UI framework for web, iOS and Android. Services tech lead Daniel Andrade Groppe said CHAOS lacked text input fields and network requests, and extending it would have delayed delivery. When a chat starts, the client fetches the CHAOS view configuration once and initializes a dataset of message events; later events (persistent user and bot messages, temporary typing indicators and quick replies, and links) update the view without refetching the configuration, and the backend can also send actions, for example returning a message, a link to view matches and an instruction to disable input when a project request is complete. Recoverable errors trigger retries, while non-recoverable failures disable input and ask the user to try again later. Yelp lists the costs (engineers had to learn both the Assistant and CHAOS, and shared implementations can still behave differently across platforms), but separating view configuration from conversation data let it reuse the Android Assistant for its Biz Ask Anything pilot. Source

4. Thoughtworks Pairs Transaction Foundation Models With Graph Neural Networks in One Shared Context Layer for Bank AI

Thoughtworks published two posts on November 6 from a Sibos 2026 breakfast it hosted with AWS and NVIDIA, arguing that bank AI pilots stall on missing context rather than model quality. Nathan Hilt separates two architectures for fraud: a transaction foundation model (TFM) trained on event sequences learns velocity, timing, merchant category and device history within one entity’s history, while a graph neural network models customers, devices, merchants and cards as connected nodes to catch coordinated fraud, such as mule rings and synthetic identity clusters, that looks normal one account at a time; neither replaces the other, and he cites Revolut’s PRAGMA (24 billion events, 64.7% better external fraud recall from one pretrained backbone) as an example. In a demo, Thoughtworks’ Brian Blanchard ran the same batch of transactions through Claude with and without transaction context, and only the grounded run returned a structured risk breakdown with a count of transactions for immediate review; AWS’s Paul Chang said 25 months of history is typically enough to start.

Rav Hayer frames the cost side: 20 teams in one bank each re-uploading the same customer and transaction data to rebuild context is an architecture problem, and the fix several institutions are converging on is one shared intelligence layer, trained once on the institution’s own history, that fraud, credit risk, payment routing and customer experience teams all draw from, with frontier models orchestrating complex reasoning and smaller fine-tuned open models handling narrow, repeatable tasks close to the sensitive data. Both posts call for governance from day one: a record of every question asked of a model and every answer, limits on what data each model or agent can access, and the ability to reconstruct a decision for a regulator. Source Source