<- all tokdocs

Atomic Skills vs Workflows: Do Agent Skills Pollute Context?

By , founder of Mavric and Apso

View on TikTok ->

Atomic is an open-source TypeScript runtime for coding agents that supports both Agent Skills and its own multi-stage workflows. A TikTok from @theneurondaily, cut from a livestream with Atomic contributor Alex Lavaee, starts from the premise that installed skills pollute the context window and recommends removing most of them. The primary sources describe a smaller cost: an unused skill puts only its name and description in context, which Anthropic estimates at about 100 tokens.

Where does the clip come from?

The clip is from The Neuron's 2026-10-08 livestream, titled "The Future of Coding Agents, According to Microsoft Research | Alex Lavaee." The newsletter that day said Lavaee would show "how to test and verify coding agents while using Atomic to build a 3D Subway Surfers-style game in the background." A companion article published on 2026-10-09 names the hosts as Corey Noles and Grant Harvey.

Lavaee's own site describes him as a "software engineer at Microsoft, working on AI/ML." The Neuron calls him a "Microsoft Research engineer" and says he built Atomic.

What is Atomic?

bastani-inc/atomic is an MIT-licensed TypeScript project from Bastani, Inc., created on 2025-10-23. The README says "Atomic is a fork of Pi" and that "Atomic implements the Agent Skills standard." As of 2026-10-19 it had 862 stars and a latest release of 0.9.39. Lavaee (GitHub user lavaman131) is its second-largest contributor.

The README describes "three top-level building blocks: workflows, skills, and specialized subagents."

Do agent skills pollute the context window?

Much less than the clip suggests. Anthropic's Agent Skills overview describes three loading levels:

  • Level 1, the name and description, loads "Always (at startup)" at "~100 tokens per Skill."
  • Level 2, the SKILL.md body, loads "When Skill is triggered" at "Under 5k tokens."
  • Level 3, bundled files, costs "None until accessed."

The page states the result directly: "until a Skill is triggered, only its name and description occupy context." The open spec at agentskills.io describes the same sequence and caps a description at 1,024 characters.

Atomic follows the same model. Its skills documentation says "only descriptions are always in context, full instructions load on-demand." The upstream Pi skills doc adds that Pi puts each skill's name, description and path in the system prompt and "does not add the full instructions."

Using Anthropic's figure, 50 installed skills would add roughly 5,000 tokens of system prompt before any of them fires. Atomic's docs give no token figure of their own. Whether that counts as pollution depends on how many skills you carry and the size of your model's context window. The README does name the general concern ("Large, mixed, or growing contexts can make attention harder") and answers it with specialized subagents that provide "isolation, focus, tool scoping, and deliberate handoffs."

Should you remove your skills and use workflows instead?

In the clip Lavaee says, "I recommend removing all of your skills that aren't like absolutely necessary and just write custom workflows for them instead." He suggests running Atomic "without your custom skills first" to see what you need.

The README is more conditional. Under "Bring your skill stack" it says, "Already have agent skills? Bring them into Atomic," and notes that "configured Claude Code or Codex skill directories can be used without rewriting them." It recommends converting only "when a skill captures a repeatable process." Its decision rule reads: "Use direct chat for small, interactive work. Use a skill or bounded subagent when the parent should stay in control. Use a workflow when a delegated job needs durable stages, retries, evidence, resumability, or approval gates."

Atomic also ships its own skills, including research-codebase, create-spec, prompt-engineer, skill-creator and tdd. The blanket advice in the clip is Lavaee's live opinion. The TikTok caption ("keep only the skills you actually need and move repeatable multi-step processes into workflows") is closer to the documented position.

Atomic workflows vs skills: what do you gain and lose?

A skill is a SKILL.md file that, per the README, "teaches the agent a procedure it loads on demand - the same format Claude Code and Codex use." A workflow is a TypeScript workflow({...}) definition whose "stage dependencies must form a directed acyclic graph," with "inputs, stages, branches, parallelism, retries, checks, artifacts, checkpoints, and human review gates."

That structure is what the clip means by "a much more powerful version of skills." The README puts it this way: "Markdown helps set context, but a model still has to follow it. An Atomic workflow runs declared stages and tools, validates configured outputs, records configured artifacts, and applies defined gates."

The clip leaves out portability. The Agent Skills site lists dozens of compatible clients, including Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and VS Code. An Atomic workflow is TypeScript written against Atomic's runtime, so converting a skill ties that process to one harness.

Can Atomic write the workflow for you?

Yes. The README says "Atomic inspects the skill and writes reviewable TypeScript" and includes the prompt to paste, which asks for "a reusable TypeScript workflow({...}) that preserves the skill's intent." The Neuron's article attributes to Lavaee a progression in which "prompts request work, skills package reusable knowledge, and workflows make repeatable process logic executable."

Bottom line

Idle skills cost a description line each, about 100 tokens by Anthropic's estimate, so the context argument for deleting them is weak. Atomic workflows add executed stages, checks and approval gates, and they lose the portability of a SKILL.md. The README's own advice is to keep your skills and convert the ones that capture a repeatable process.

Sources

Published October 16, 2026. Based on a TikTok by @theneurondaily. Full video notes (markdown)