Dated observations. First person. What changed, what doesn't work, what's becoming standard.
Across 267 tracked sessions, 62% last under 10 minutes. Long sessions (>1h) account for only 15% of the total. The micro-session is the real productive pattern — not tool underuse.
Across 128 auto-scored Claude Code sessions, 68 had a "wrong_approach" friction — 53%. The agent's first instinct is wrong half the time. This isn't a problem to fix, it's a pattern to integrate into the workflow.
Kilo Code CLI is a fork of OpenCode with Kilo Gateway on top. In practice, the TUI is identical, custom commands work the same, and migration took 15 minutes. The real difference: free models via the gateway change the workflow economics.
Three patterns that naturally emerged in a daily agentic workflow: automated morning briefing, versioned custom commands in Git, and parallel sessions in tmux splits. Each came from a real friction, not from theory.
The tool wasn't the problem, the model was. Kilo Code + Mistral looked perfect on paper. In practice, Mistral nearly killed adoption. Pivot to OVH AI Gateway + Gemini. It works now.
The bigger the context window, the later context rot hits — it doesn't go away. The real solution is structuring your work into short sessions with verifiable deliverables. It's an organizational discipline, not a hardware problem to solve.
On mechanical, repetitive PHP/TypeScript tasks, budget models (MiniMax 2.5 via Kilo Gateway, Gemini Flash) produce results indistinguishable from Claude Sonnet. On complex reasoning tasks, Claude maintains a real edge. The takeaway isn't "use the cheapest" — it's "use the right model for each type of task."
Claude Code is an excellent starting point — zero config, immediate quality. But Anthropic lock-in prevents optimizing costs per task, and restricting the Claude Code subscription in third-party tools forced the migration. OpenCode then Kilo Code CLI (its fork with free gateway) solve both without sacrificing quality.