# AI coding goes from generation to verification > Source: > Published: 2026-11-06 17:24:06+00:00 ICYMI AI: The latest AI trends, delivered to Dev.to In the early days of AI-assisted code writing, having AI generate code meant providing suggestions you’d review. Yet emerging research demonstrates that newer AI systems are checking, testing, and improving their own work at scale. Are you starting to wonder where your role fits in this evolving workflow? Axiom Math [directed swarms of agents](https://x.com/axiommathai/status/2115089794867040584) for 24 days to produce roughly 43.1 million lines of Lean code, the first fully computer-checked proof of the Classification of Finite Simple Groups— a big step toward finding AI verification useful for evaluating AI-generated code. [OpenAI](https://openai.com/index/building-scientific-understanding-with-academic-communities/?utm_source=therundownai.beehiiv.com&utm_medium=newsletter&utm_campaign=rundown-1027&_bhlid=fcde9531abe55e678e0fc7523c7cc236571ff9b0) and [Anthropic](https://github.com/anthropics/formal-math/tree/main/eth) each announced model-found results on the Exponential Time Hypothesis with Lean-verified proofs. Now, the same coding loop that developers can run daily to write, check, and iterate can sustain weeks of unsupervised work and tackle more complex coding issues. Production engineering is following the same recipe: For developers, the shift comes down to moving away from generating code and toward defining verifiers: tests, type systems, replay harnesses, and acceptance criteria that an agent can iterate against. Even though Claude wrote [90% of Anthropic’s code](https://claude.dev/blog/solving-code-review-at-anthropic/) in September, the process still required a heavy human hand in deciding what to build, risky code and architecture review, testing outputs, and owning outcomes. Using AI to verify its own generated code is seeing new advancements—but still needs a developer’s knowledge to double-check outputs and build the verification workflows to effectively produce usable results. **The open-source engine behind a growing share of AI inference just hit production-ready**. With [llm-d 1.0](https://newsroom.ibm.com/2026-10-20-llm-d-1-0-advances-open,-cost-efficient-ai-inference-for-production-scale), now a CNCF project backed by Red Hat, IBM, Google Cloud, CoreWeave, and NVIDIA, the contributors cite demonstrations of 5 to 10 times lower costs for context-heavy agentic workloads and add multi-tenant quality of service and zero-downtime upgrades. DigitalOcean’s Salman Paracha, SVP of AI Engineering, shared how DigitalOcean relies on llm-d as foundational distributed inference software for Serverless Inference and Managed Agents, serving trillions of tokens. **Writing more code isn’t the same as shipping more software**. A [Harvard working paper](https://fion.ac/jellyfish.pdf) analyzing about 300 million engineering events at 718 firms found that after adopting fully agentic coding tools, lines of code rose 30% and pull requests rose 23%, yet resolved issues and epics showed no statistically significant gain. The likely culprit is review: the average time to review a pull request rose 49%, so the extra code piled up at the review stage. **AI assistants are moving from answering questions to using a computer for you**. [Claude Duet](https://techcrunch.com/2026/10/22/claude-duet-app-is-anthropics-bet-on-agents/), now in research preview, runs in its own cloud environment via a browser, integrates with 6,000+ connectors and plugins, and requires approval before payments or other sensitive steps. It joins Grok Bot, Dots, and Muse, and our [side-by-side](https://www.digitalocean.com/resources/articles/grok-bot-vs-dots-vs-muse) comparison shows how they differ on connectors, data use, and cost. Video production used to require serious budgets, professional crews, and weeks of post-production. Higgsfield AI is changing that by putting high-quality AI video creation in the hands of social media marketers, content creators, and brands. To provide professional-grade video to its users for \$15/month, the company needed a reliable, cost-effective AI cloud provider. When neo-clouds couldn’t meet production workloads, Higgsfield AI turned to DigitalOcean and set up infrastructure that reliably supports its users’ video creation. **DigitalOcean Serverless Inference latency: p50, p95 and p99 TTFT for GLM-5.3, DeepSeek V3.2, Gemma 4 and Llama 4 Maverick (3,456 requests, 48 hours)** We timed four models on DigitalOcean Serverless Inference for 48 hours. Across 3,456 requests, DeepSeek V3.2 started replying in 1.35 seconds at p50 but 13.7 seconds at p99, while GLM-5.3 streamed fastest at 59 tokens per second. Get the p50, p95, and p99 TTFT and tokens-per-second results for all four models, plus the published script and raw data to rerun the test yourself. **Agents in Production, Part 3: Let Your AI Support Agent Send Replies, Without Giving It the Credential** In two live runs covering 335 tickets, an approval rule on send_reply never fired because the agent only ever called action_invoke, and the agent marked 78 drafts as routine, which the support board then refused. Learn how to keep the credential out of the agent with Action Gateway's credential brokering, close the network with an egress allowlist, and enforce the send rule where it actually holds. **How Do I Monitor and Trace AI Agents Running in Production?** We ran a test feeding one bad tool result across multiple sessions. Every tool call succeeded, even though the answer was still wrong, and an automated debugger, given only a strict application log, found the bad call in 0 of 15 tries. See what to record instead: with every tool result and a check on each record, the debugger found it in 15 of 15, and the article includes a capture checklist and decision table you can copy. **Sandboxing LLM-Generated Code Needs a Different Threat Model** The isolation ladder from bare process to microVM was built for adversaries, but across 92 traced agent sessions, the median session that executed code sat idle 94.5% of its duration. Walk through a 1,500-row failure battery across five isolation tiers and a decision framework for which rung your workload belongs on. **Now Generally Available: Prompt Caching for DigitalOcean Serverless Inference** Cache affinity now makes your hit rate reliable rather than accidental, lifting flagship models to 85–95%+, so the up-to-90% discount on cached input and up to 80% faster time-to-first-token actually show up on your bill. A new dashboard tracks hit rate and savings per model, and prompt_cache_key lets you control how traffic is grouped or isolated.