Gemini 3 Flash vs Pro for coding
Gemini 3 Flash scores 78.0% on SWE-bench Verified vs 76.2% for Pro, at 75% lower input price. Why distillation explains it, and the trade-offs.
Series
Exploring AI-assisted development and agentic tools
I ship code with an agent doing most of the typing now, and I write this series because the tools change faster than anyone's advice about them. Most agentic coding content is either a vendor pitch or a doom post. I want the third thing: what actually changes how I work, tested against my own repos, not a demo.
The pattern across these posts is that the interesting decisions live outside the model. An identity file changes defaults more than a better prompt does. A hook that reinstalls rules after compaction matters more than a longer system message. Orchestrating five agents is a state problem, not a prompting problem, and I say so even while arguing that most teams reach for the heavy version too early. I also cover the boring stuff that gets ignored: agents authenticating as me instead of as themselves, and why cheaper, smaller models already beat the expensive ones on the benchmarks that matter for coding.
None of this is theory. Every post here comes from a tool I installed, read the source of, or ran against real work.
Start with the mechanism underneath everything else here: an identity file changes an agent's defaults, which does more than a better-worded prompt.
The same idea taken further — a rule system that assumes suggestions do nothing, with a hook that reinstalls itself after compaction.
The day-to-day workflow this enables: a spec-driven loop built to survive context rot instead of hoping a bigger window fixes it.
What happens once that loop is reliable enough to run unattended overnight.
What happens once one agent isn't enough — coordinating several turns out to be a state problem, not a prompting problem.
The operational risk that shows up once agents are actually running: none of them have their own identity, and the audit log knows it.
Gemini 3 Flash scores 78.0% on SWE-bench Verified vs 76.2% for Pro, at 75% lower input price. Why distillation explains it, and the trade-offs.
Claude Code has the best agentic experience but costs ~$1k/month for heavy use. I built a Claude Skills-compatible system for Antigravity that makes AI coding preferences portable across platforms.
Cross-platform frameworks solved expensive labor. When AI makes labor cheap, the abstraction overhead remains but the savings vanish. Native is coming back.