The hard part of coding agents isn't the code — it's the 4 jobs around every PR
Arindam_1729 · x · 2026-10-31
Arindam argues coding agents are already good at the code itself — pointing at an issue, reading the repo, editing files, running tests. The hard part is the four jobs you still do around every PR: picking which issue is safe to hand over, opening the PR, pasting CI failures back in, and deciding when to give up.
He breaks down how @superplanehq turns this into a factory line: backlog > implement > verify > done:
- Intake scores each issue before an agent starts
- Verify runs your real checks first, feeds failures back as short excerpts, and stops after a set number of retries so it can't burn tokens forever
- The author is candid: it's only as good as your tests — weak CI, weak guardrails
Full setup guide linked in the article.
More from coding & agent
- Three tips to make your model stop overcautious actions and converge faster — tokenbender · 2026-11-01
- Lean creator Leo de Moura on why AI-generated code needs formal proofs — bigdata · 2026-11-01
- Maybe we're keeping humans in the wrong part of the AI coding loop — RoyPartha04 · 2026-11-01
- Context isn't the agent bottleneck — enterprise knowledge is an untrusted mess, argues Google researcher — _AustinCalvert_ · 2026-11-01
- Engineer: AI PCB design doubled my workload — fixing its errors slower than doing it myself — cneuralnetwork · 2026-11-01
- Claude Code drives an oscilloscope over the network to automate waveform debugging — cneuralnetwork · 2026-11-01