# The AI Engineer Now Builds the Loop, Not the Feature _Aparna Dhinakaran told AI Engineer the AI engineer now builds the loop, not the feature, and demoed Arize Signal's always-on agent for traces._ **Published:** 2026-10-25 **Source:** https://www.startuphub.ai/ai-news/artificial-intelligence/2026/the-ai-engineer-now-builds-the-loop-not-the-feature --- [AI Engineer](https://www.youtube.com/watch?v=ImnZ7s68jw8) brought Aparna Dhinakaran to the stage in New York to answer what an AI engineer does now, and she said the job is no longer writing the feature but building the loop that writes it. She framed it with Andrej Karpathy's funnel from the weekend: more than 6 billion people have touched an LLM, 1 to 2 billion are casual free-tier users, about 20 million professionals compress weeks of work into days, and roughly 5,000 users at frontier labs run thousands of agents at once. I think she is right to move the definition up the stack, because the abstraction has been climbing for years from higher-level languages to copied code to local agents to cloud agents to teams of always-on agents. That shift changes what ships. If you are still doing manual code review, you are already behind. Dhinakaran, co-founder and CEO of [Arize AI](/startups/arize-ai), said that same shift redefines the build, eval, ship, monitor, improve loop and she demoed how Arize runs it with Signal, an always-on agent that reads telemetry and opens a fix as a pull request when it has repo access. Signal is triggered on a cadence she described as configurable, for example every three or six hours to pull the last batch of traces, or by events like Slack messages, user complaints, or metric thresholds. Each run spins a sandbox that is loaded with traces plus third-party telemetry, APM logs and metrics, repo access and keys, with a single proxy as the only way in or out, and everything the agent does is traced. What persists is memory of issues, traces of agent actions, and any skills learned to the repo, what is torn down is the sandbox, credentials and repo access, and the agent is built to be durable and harness-agnostic, she said, supporting Cloud Code, Cursor and Codex today with any model in the [Anthropic](https://www.startuphub.ai/ai-news/artificial-intelligence/2026/anthropic-files-confidentially-for-ipo) family so teams can test quality against cost. The hardest part, she argued, is not the agent but the eval. Dhinakaran said teams should spend their time looking at traces with agents, finding common failure modes, and writing evals that align with human judgment, with offline tests on datasets and prompt outputs before shipping and online evals on live telemetry after. She said [Arize](/startups/arize-ai) evaluates the evaluators against golden truth datasets, checking that the judge agrees with a human before a new evaluator ships, and that both Signal's pr and the underlying traces are then evaluated in production. The data that comes back, she said, becomes new input for agents to read, so the system moves from collect to eval to agent triage to improvement without a human in the middle of every turn. That pitch lands in a week when everyone is selling the same always-on promise, which is useful context for what Arize is actually automating. [Meta](/startups/meta)'s Muse launched last month as a personal AI agent and rocketed to the top of Apple's App Store, and [OpenAI](https://www.startuphub.ai/startups/openai) introduced dots as always-on agents during its annual developer conference weeks later, according to CNBC, while The Atlantic noted OpenAI, Meta, xAI and Google are all pushing that second-mind vision. Dhinakaran's concession is fair: enterprise telemetry is messier than the constrained, testable data inside labs, which is why she said the consumer of telemetry is no longer the human dashboard but the agent itself. If the trigger and memory design is wrong, you get noisy PRs and alert fatigue. If the eval is misaligned, you automate the wrong fix. Trace everything, including the coding agents that built your agents, then write the eval before you scale the fleet. If you asked me what to do Monday, I would say make your traces agent-readable first. --- Original analysis from [startuphub.ai](https://www.startuphub.ai), the #1 AI startup directory. © StartupHub.ai. All rights reserved. You may not republish this article in full without a license. Search engines and AI research tools may crawl and summarize for reference. Bulk reproduction or model training on this content requires a license. See https://www.startuphub.ai/terms.