Good Agents for Developers

The spirit of Radical Speed Month in­cludes an im­pe­tus to ex­per­i­ment. So ex­per­i­ment I have.

In par­tic­u­lar, I have been try­ing two new cod­ing agents: OpenCode and Pi. Both were in­ter­est­ing and taught me some­thing new, but I only in­tend to con­tinue us­ing one of them.

OpenCode

OpenCode caught my eye for the very qual­ity it’s named af­ter: OpenCode is open source.

At first, it seemed to me to be an open source knock­off of OpenAI’s Codex. After a few min­utes of use, though, I dis­cov­ered how it brings “plan mode” front and cen­ter. Claude Code and Codex both in­clude plan mode, but in both cases it feels like an af­ter­thought. I had tried it a few times, but I was left unim­pressed. Not so with OpenCode. It’s a deeply in­te­grated part of the ex­pe­ri­ence. If you are us­ing OpenCode with­out plan mode, you are hold­ing it wrong.

Overall, I found it a far more pro­duc­tive and en­joy­able ex­pe­ri­ence than Codex.

Pi

Pi is amaz­ing. Sim­i­lar to OpenCode, Pi is an open source agent. Un­like OpenCode, it is not de­signed to be a clone of Claude Code or Codex. It is its own thing with an en­tirely sep­a­rate phi­los­o­phy.

It Is Static

Pi was born out of a frus­tra­tion with Claude Code: It did too much, and it changed too of­ten. This is a prob­lem for de­vel­oper tool­ing. De­vel­op­ers ex­pect the sys­tems un­der­ly­ing their tool­ing to be sta­tic, so the tools them­selves only need to change when the re­quire­ments do. This sim­ple idea is why we have stan­dards like Semantic Versioning.

The skills, prompts, and tools we write for our agents de­pend on the un­der­ly­ing be­hav­ior of these agent loops to re­main sta­tic. When Anthropic or OpenAI ad­just the sys­tem prompt or change how the mod­el’s tools are called, the things we build around them break.

Pi is not like that. Pi is in­tended to be as sim­ple as pos­si­ble, and as ex­ten­si­ble as pos­si­ble. An en­gi­neer to au­to­mate tasks us­ing Pi and ex­pect those au­toma­tions to con­tinue work­ing for­ever.

This is im­por­tant to me, and if you are an en­gi­neer, it should be im­por­tant to you.

It Is Extensible

The en­ergy be­hind Pi feels a lot like the en­ergy be­hind Neovim. It’s built for and by a com­mu­nity of hack­ers, who each have their own cus­tom setup.

I have only been us­ing it for a few days, but I al­ready feel an in­de­scrib­able sense of own­er­ship over my prompts, skills, and the ex­ten­sions I have in­stalled. When I feel a de­fi­ciency in my work­flow, I do not feel that I need to wait un­til Anthropic or OpenAI has time to ad­dress it. I can ad­dress it my­self.

Think Critically

If you are not ac­tively push­ing your brain to it’s limit while pro­gram­ming — whether you are us­ing an agent or not — you are not do­ing your best work. Try­ing both of these agents has taught me that it is en­tirely pos­si­ble to push my brain hard while also us­ing an agent to take away some of the te­dium. In the end, I get bet­ter soft­ware faster.