November 2026 Summaries

Outdated inline TODO and FIXME comments can misdirect coding agents because agents may interpret obsolete bug references as actionable context, increasing the risk of unnecessary investigation, scope expansion, and incorrect implementation decisions. Analysis of millions of agent sessions found that 13% exposed agents to bug-linked comments, with 6% prompting follow-up action; 83% of these actions were not requested by developers. Among unprompted detours, 41% involved closed, fixed, obsolete, duplicate, or restricted tickets, while 57% involved still-open bugs that agents investigated or modified despite being outside the requested task. After controlling for multiple task and codebase variables, these detours were associated with 70% more steps, 77% more tokens, and 105% more wait time per session. To reduce this source of distraction, automated agents scanned codebases for references to resolved or obsolete bugs, reviewed their historical context to determine whether comments remained necessary, and submitted confirmed removals for review, resulting in thousands of deleted references and more than 95% reviewer acceptance over six months.

Autonomous coding-agent economics are better assessed by cost per successfully verified task than by token price alone, because retries, tool calls, reasoning passes, and failures can substantially affect total spending. A 50-question benchmark found that a lightweight classifier, TypeSafe’s Jev, could accurately route short coding tasks between Gemini Flash-Lite and Gemini 3.8 Flash with low thinking, matching the accuracy of using Flash low-thinking for every task while potentially reducing cost and response time. The study reported that broad difficulty labels and fixed regex or prompt-length heuristics performed poorly on unfamiliar tasks, whereas a narrowly defined binary classification question about whether a task required multiple reasoning steps achieved stronger results. It also found that lowering output-token limits did not reduce reasoning-token use and instead caused failures on harder tasks, while explicitly setting a lower thinking level was more effective. The results are limited by the small, single-run benchmark, and the discussion notes that routing may undermine model-specific context-cache savings for long, repository-heavy interactions, making single-model use preferable in those cases. The central recommendation is to measure real workloads, begin with a lightweight model configuration, route only suitable standalone tasks through a classifier, and evaluate approaches by the cost of correct outcomes rather than token rates.