Agentic Coding 2.0: Eval-Driven Loops and the New Daily Grind

Agentic coding finally feels less like a buzzword and more like the daily bread for anyone building with Claude, GPT-5, or the new Gemini variants. The old dance—write prompt, hope for gold, pray it doesn't hallucinate—has been replaced by what I call the eval-driven loop: plan, act, verify, fix, repeat. It's noisy, sometimes brittle, but the productivity boost is real.

Spec-Driven Prompts: No More Guesswork

The game changer? Spec-driven prompting. Instead of handwavy instructions, you feed the model a bulletproof spec—types, edge cases, expected behaviors. It’s contract programming for LLMs. With Claude Code and similar tools, I routinely hand the model a problem, a test suite, and a type signature, then let it iterate (via loop engineering) until the test harness goes green. This isn’t magic—it's engineering: codifying what "done" means, and letting the machine grind until it gets there.

MCP Tool Servers & Real-World Glue

Most devs I know have MCP-style tool servers glued into their stack. These act as a bridge between LLM reasoning and "real" APIs, letting agentic workflows fetch docs, call DBs, or spin up ephemeral cloud instances. The setup isn’t trivial, but once the loop is tight, you get agents that can reason, act, and self-correct in production—no more toy demos.

Why This Matters

As eval-driven workflows get tighter, devs are spending less time on prompt engineering and more on orchestrating robust loops and monitoring real-world outcomes. AI pair-programming isn’t just chat; it’s a continuous harness, where the model proposes, you review, and tests arbitrate. It’s humbling, but it’s the only way to scale trust and velocity.