How Anthropic Makes Claude More Reliable

Living at the edge of AI is bittersweet. You can spend weeks building a workaround to a problem only for a frontier lab to swoop in and solve it for you in a more elegant, reliable way. Today, senior applied AI engineer Nityesh Agarwal explains how Anthropic’s dynamic workflows feature made his elaborate Claude setup look clumsy in retrospect, the Every team shares which corners of the AI frontier they’ve given themselves permission to ignore, and executive operations manager Jalaiyah Bolden walks through her step-by-step process for turning a Slack bot into a reliable coworker.

Every is off tomorrow for Juneteenth; we’ll be back Sunday. Was this newsletter forwarded to you? Sign up to get it in your inbox.

Mini-Vibe Check: Dynamic Workflows

A closer look at how Claude Code coordinates multiple agents

When senior applied AI engineer Nityesh Agarwal built Every’s AI project manager Claudie, he spent days figuring out how to get around the model’s limited context window, or the cap on how much text an LLM can process at once—and the reason Claudie kept dropping key details. His solution: one coordinating agent that delegated tasks to fleets of subagents, which gathered data, made updates, and communicated with one another via local markdown files. The process was “a little bit hacky,” Nityesh says, but it worked.

If he were to build Claudie today, he could just use dynamic workflows, Anthropic’s feature for orchestrating large, multi-agent Claude Code tasks. Instead of deciding each step on the fly, Claude writes a reusable script that coordinates the work. It can assign tasks to many subagents and have them check each other’s work before reporting back the results.

Before dynamic workflows, trying to get Claude to reliably spawn reviewer agents was a persistent headache. Anxious about token spend, the model “would sometimes try to merge it all into one subagent,” Nityesh says, dragging down the quality of the results. Increasingly dramatic directives not to do this often went unheeded. Now, if you tell Claude you want three verifier subagents with dynamic workflows, Claude will write a script that generates three subagents every time.

Nityesh is grateful for the new feature, but watching weeks of work get negated by a single release was also disheartening. “I spent so many weeks building that other thing. Now it’s useless,” he says.

“But that’s the cost of being at the frontier,” he continues. “You need to be ahead of everybody else, and sometimes that means you need to throw away your past work.”

A dynamic workflows case study. For Spiral’s redesign, senior designer Daniel Rodrigues sent the writing app’s general manager Marcus Moretti a giant Figma file.

Marcus needed to convert the file into code. He did a pass in Claude Code, but the result had numerous errors. Before dynamic workflows, he would have flagged the mistakes in batches for Claude Code to fix—a repetitive, frustrating process.

Instead, Marcus asked Claude Code to set up a dynamic workflow that would review the Figma file section by section, extract all assets and design details, turn them into code, and check the results against the original file.

The Figma file had 11 sections, so Claude spun up 11 tasks, each with dedicated subagents. After running for a couple of hours, “it was not perfect,” Marcus says, but “it saved me a whole bunch of time.” Before dynamic workflows, each of the reviewer subagents would have been Marcus himself.