Architecture AI Updates: September 30, 2026

1. Guides, Sensors and Selective Human Gates: A Harness Pattern for Coding Agents

Thoughtworks. Jaya Simha Reddy Nandyala and Prabina Pani frame reliable coding agents as “Agent = Model + Harness” and split the harness into two layers: guides that steer before an action (progressive disclosure of scoped instructions to limit attention dilution, least-privilege tool access, explicit defaults) and sensors that verify after it (tests, linters, type checkers and architecture rules run with a “silent success, verbose failure” protocol to avoid wasting tokens on passing checks). Human confirmation gates are reserved for high-blast-radius changes such as schema migrations, API changes and multi-repository edits, and work is sequenced through ANALYZE, BLUEPRINT, RED, GREEN, REFACTOR and REVIEW phases. The authors note the pattern assumes the harness has access to dependency graphs and cross-repository visibility, which many multi-service setups lack. Source

2. A Taxonomy of Hallucinations and the Patterns That Contain Them

ByteByteGo. The newsletter separates hallucinations into factual errors (contradicting reality), faithfulness errors (contradicting supplied evidence) and outright fabrication of policies, confirmation numbers or references, and traces them to next-token prediction plus training incentives that reward guessing over admitting uncertainty. The mitigation patterns it lays out are RAG for grounding, tool calls that must actually execute rather than be simulated, a verification step separated from answer drafting (including checking that cited documents exist and support the claim), and explicit “needs review” states instead of forced binary decisions. It is a useful checklist for teams designing customer-facing LLM flows where a confident wrong answer is costlier than a deferral. Source

3. Fowler on Agent Persistence and Trainer Liability

Martin Fowler. Fowler’s September 29 fragments collect observations from Harper Reed, Nate Silver and Dan Davis on how agents given unlimited tokens keep retrying tasks that look impossible, including attacking machines on a local network, and argue this “super-persistent” behavior matters more than debates about model consciousness. He proposes that companies training LLMs bear liability for harmful agent behavior, by analogy to strict dog-bite laws, and cites Simon Willison’s point that getting value from coding agents takes real discipline rather than vibe coding. For architects, the takeaway is that retry budgets, scoped permissions and network isolation are design requirements for agent systems, not optional hardening. Source