12 articles
Max Heckel is the founding engineer and CTO of Ariso. Before starting Ariso, he worked at Google, McGraw Hill, JupiterOne, and created SciSummary.
LinkedIn ↗We moved bounded classification decisions in Ari from general-purpose LLM calls to TypeSafe's Jev, at roughly one-twentieth the cost and 10–15× the speed. Inside the implementation: typed questions, runtime validation, explicit uncertainty policies, and LLM fallbacks for the work that still needs them.
At Ariso, we each review tens of pull requests a day. A GitHub Action prepares the handoff: when a draft becomes ready for review, Claude Code assesses risks, blockers, regressions, deleted behavior, and specific human checks. Here is how it works, why the definition of a blocker matters, and the complete workflow YAML.
AI can write code quickly, but shipping reliable software remains a human endeavor. The operating model I've built at Ariso: sprint planning becomes a dashboard of to-dos, I rewrite each in my own words inside Ivan, Ivan opens draft PRs, and custom widgets surface the review queue. Then ivan review preps me as a human reviewer, turbo-hot-cache warms the branches, and I work through the diffs — reading, running, and testing each one before moving it from draft to ready for review. Machines produce and organize the work; a person verifies it before it moves forward.
A hot take sparked by a Reddit thread about AI burnout: junior engineers should still learn data structures, architecture, and system design, because none of it is tied to a specific language. Syntax was tied to a language from the start, which is exactly why a coding agent already does it better than any human ever will.
Every week an engineer I respect tells me they're worried a coding agent is coming for their job. Engineers spend about 30% of their time writing code, and the more senior you get, the less you write. The agents took the typing. Here's how we direct and judge all that new capacity at Ariso, from Slack-triggered fixes to ivan review keeping humans on top of the diff flood.
I opened 80 pull requests in a two-week sprint by assigning tasks to an agent and walking away from my machine. The tool is called Ivan, it's free and open source, and it runs natively in GitHub Actions. Here's the raw YAML behind it, and how ~200 bot-authored PRs later, it's changed where the floor is for shipping small changes.
We mentioned a bug in a meeting on Tuesday. This morning, Claude Code suggested fixing it the moment I opened my terminal — and twenty minutes later there was a PR. Here's how ari-hooks wires Ari into Claude Code, and why we don't use a ticketing system anymore.
If your standup is still a tour of what everyone — or everyone's coding agent — did yesterday, you're doing it wrong. Ari already knows what happened. Standup isn't the memory layer anymore; it's the steering layer. Here's how we run it at Ariso.
This isn't about autocomplete anymore, or a slightly smarter Clippy. We're talking about a bot that took a feature request from a chat message, understood the business context, scanned the entire codebase, and opened a production-ready Pull Request. It did it in 20 minutes. While we were getting coffee, the software built itself. This isn't the future. This is Tuesday.
Managing a coding CEO isn't about stopping them from coding. It's about building a padded room where they can do cool karate moves without breaking the furniture.
Coding agents aren’t going to steal your job—but engineers who learn how to work with them will absolutely outpace those who don’t. At our company, we don’t treat AI as a novelty or a replacement for thinking. We treat it like a junior engineering team that never sleeps. Every night, we hand off a list of tasks to our coding agents. By morning, we’re staring at a stack of pull requests—some useful, some disposable, all cheap to generate.
Hidden context messages keep user-facing LLM chats clean while giving the model rich, per-message instructions—turning conversations into a predictable, tool‑driven workflow engine with a tiny amount of code.
