No-Code Agentist

Most people think AI adoption = more chat messages. Wrong. New data on 100K+ users shows something different: People aren't talking to AI anymore. They're delegating entire workdays to it. 🧵 What the Codex data actually reveals: For the past 2 years, we've measured AI adoption the wrong way: • Monthly active users • Messages sent • Queries per day But these metrics miss the entire story. The real shift isn't conversational → more conversational. It's conversational → agentic. Users went from asking questions to handing off multi-step tasks that run for hours while they do other work. Inside OpenAI, 99.8% of output tokens now come from agentic workflows. Not chat. Not back-and-forth. Delegation. These are tasks that would take a human 8+ hours to complete. Think of it this way: Old: "Can you help me debug this function?" New: "Build the entire feature, run tests, integrate with API, document it—I'll check back in 3 hours." That's the gap. The most sophisticated users run 5-10 agents concurrently. They're not using AI as an assistant. They're using it as a distributed workforce. Parallel delegation is the new normal. Where did this start? Software engineering. Why? Two reasons: Tasks are verifiable (code runs or it doesn't) Engineers were already thinking in modular workflows But it's spreading fast: • Legal teams: Contract review, research memos • Operations: Process docs, data pipelines • Research: Literature synthesis, analysis Anywhere verification costs are manageable. Here's the leverage point most orgs miss: Reusable "skills" (workflow templates). Users who build custom skills for their team see 3-5x higher adoption rates. Why? They capture org-specific context once, reuse forever. Example skill: "Take this API spec → generate client library → write tests → create integration docs → submit PR" One skill. Invoked 47 times across the team in a quarter. That's 376 human-days of work, delegated. The constraint is no longer: "Is the model smart enough?" It's: "Can I clearly specify what I want?" "Can I verify the output efficiently?" Management becomes the bottleneck. Hot take: ChatGPT is already legacy tech. The product isn't the chatbot. It's the agent runtime underneath. Output token growth for heavy users: • Individual users: 10x in 6 months • Organizational users: 15-25x • OpenAI internal: 50x Why the gap? Organizational friction. What creates friction: • Security policies (data access) • Lack of internal training • Verification overhead • Missing workflow redesign Remove these → output explodes. The study tracked 3 groups: Individual users: Fastest to adopt, lowest complexity Organizations: Slower start, higher ceiling once skills institutionalize OpenAI internal: Asymptotic—nearly all work is agentic This is your adoption roadmap. New KPI to track: % of output tokens from agentic tasks Not "messages sent." Not "active users." Output tokens = actual delegated work. Another hot take: Software engineers will be the first middle-management class partially automated by tools they built. The irony is thick. If you're a team lead, start here: Pick your top 3 repetitive workflows Build one "skill" for each Train 5 people to use them Measure: skill invocations per week Time to value: 4-6 weeks. OpenAI's internal data shows the ceiling when all friction is removed: 10-50x output growth. That's not hype. That's measured reality in an optimized environment. The shift from chat to agentic AI is already happening. The question isn't "if." It's "how fast can your org adapt?" Bottleneck = delegation skill + verification infrastructure. Who's already making this shift? 👇