What the Developer with the #1 AI Token Usage Learned from 148,000 Conversations

The conclusion reached after over 148,000 conversations with AI over 53 days is clear. The performance gap comes not from how much you use AI, but from how you make AI work.

The developer's role is shifting from someone who writes code directly to an orchestrator. The center of actual work has also moved from implementation to decomposition, planning, verification, and recording, and this structure determined the quality of the output.

Quantifying the usage log makes the flow clearer.

• 940 directly initiated sessions
• 2,594 derived sub-agents
• 12,632 directly written messages
• 129,537 AI-generated messages
• 177,590 tool calls
• A delegation structure where AI takes about 10 actions per 1 input

AI usage levels can also be organized into 5 stages.

• Level 1: Q&A
• Level 2: Code generation
• Level 3: Tool connection
• Level 4: Orchestration
• Level 5: Multi-agent autonomy

The core was at Level 4. Even with the same Claude, the same GPT, results differ structurally depending on where they are applied.

Actual work ran through 6 stages.

• Explore: Thoroughly explore domain knowledge and the codebase
• Plan: Specify architecture and completion criteria
• Decompose: Break work into parallelizable units and assign agents
• Implement: Delegate and execute in parallel
• Verify: Verify through tests, Playwright, and API calls
• Record: Accumulate decisions and knowledge

If even one of these was missing, quality wavered. Implementing without exploration missed conventions, starting without planning caused direction to drift, and finishing without verification let errors pile up.

AI's failures also weren't random but followed repeating patterns. The observed problems can be organized into 5 types: cause reduction, hypothesis fixation, local optimization, environment blind spots, and exploration interruption. Ultimately, the root of failure lay in context quality, which meant exploration needed to be taken deeper and wider.

Analysis of tool calls converged on a read:write ratio of 10:1. Accurate exploration came before fast generation, and including existing code conventions, business rules, past decisions, and up-to-date documentation together reduced rework. Parallel exploration and securing high-quality context were the criteria that determined success or failure in AI collaboration.