Stop Writing Code, Start Writing Docs

For this section of The New Stack Agents, TNS Publisher Alex Williams and I talked to Keith Ballinger, nan vice president and wide head of nan Google Cloud Platform Developer Experience group.

Ballinger is 1 of those uncommon executives who still often codes and who is besides heavy into building and utilizing nan latest agentic coding tools. Unsurprisingly, we spent immoderate clip talking astir Gemini CLI, Google’s nonstop competitor to products for illustration Claude Code and Gemini’s Codex, and 1 of nan cardinal products Ballinger is presently responsible for.

But what’s possibly moreover much absorbing is his return connected really to get nan astir retired of these agentic tools, which Ballinger believes are still successful “their first inning.” In his view, it’s important to slow down to velocity up nan process: developers should constitute clear guides for nan agents, attraction connected architecture (and archive it) and create a task plan.

Too often, he said, developers still effort a one-shot approach, but that doesn’t supply nan models pinch capable accusation astir nan developer’s intent and nan wide discourse of nan problem they are meant to solve.

Ballinger besides believes that successful nan agelong run, we whitethorn get to a spot wherever nan mostly of developers won’t request to look astatine nan codification astatine all. The history of programming languages has been 1 of accrued abstraction. The adjacent abstraction layer, he told us, “is fundamentally my intent — what I want, and really that comes together to make an experience, to make an application.”

He acknowledged that this is simply a spot arguable and that he mostly has a “very optimistic sentiment that’s astir apt successful nan minority.”

Check retired nan afloat speech above, including a chat of Keith’s profession astatine Microsoft, Xamarin and GitHub, his Conductor model for guiding agents, and why awesome programmers are besides awesome writers. You tin besides spot much episodes of Agents by subscribing to our podcast connected YouTube aliases different podcast platforms.

What follows are immoderate of nan highlights of our conversation, lightly edited for clarity.

The Early Days of AI Coding

Alex Williams: I was asking astir really Keith, you really started to usage AI. You were saying earlier our recording, that a batch of group conscionable don’t cognize really to usage it, and they’re utilizing it successful a very single-shot benignant of way. What person you learned astir that now, and show america astir those early days erstwhile you were first getting started?

Keith Ballinger: Yeah, it’s interesting. While I’ve ever been progressive successful developer devices and technologies, arsenic a teen I was an email penpal of Marvin Minsky. So I had nan AI bug very early successful my life. I put together a hack squad successful 2012 aliases thing for Bayes Hack, which did instrumentality learning for nonprofits. So I’ve ever been really willing successful that world. [Editors note: Bayes Hack was nan hackathon of Bayes Impact.]

Then astatine GitHub, we started building Copilot. The thought location was that OpenAI had travel on and yet had a beautiful bully exemplary that, funny enough, they called Codex astatine nan time, and it could do coding tasks beautiful good if you managed it carefully. The very first point we did was purely codification completion. There was chat soon after. And interestingly, I deliberation starting pinch codification completion, starting pinch thing very interactive, was very useful successful knowing really group would usage AI and getting them comfortable a small spot pinch it.

Then group started utilizing nan chat. Some group spell to ChatGPT aliases Gemini, moreover today, to do AI coding. Then, of course, agents came along, right? Where each of a sudden, nan exemplary was doing things connected your behalf pinch instrumentality usage. That was a really absorbing maturation successful our industry.

I benignant of followed on connected it. I retrieve possibly mid-last year, tool-using models started really taking off. You deliberation astir group for illustration Cursor, each of a abrupt they were capable to make a dent because they weren’t conscionable codification completion and they weren’t conscionable chat. All of a sudden, they were doing these semi-autonomous tasks for you pinch tools.

It was funny, because up until then, for for illustration a twelvemonth anterior to that, I would conscionable usage chat. So I would spell to Gemini — we called it Bard astatine nan time— and I would show it what I wanted to do, and past it would travel backmost and springiness maine code, and I would return that codification and I would paste it into my project.

But it would besides show maine to do different things, for illustration tally a trial and springiness it nan output of nan test. I was fundamentally acting arsenic nan instrumentality for Gemini, for nan exemplary itself. And past erstwhile that was automated, each of a abrupt you could do things truthful overmuch much quickly, and it really expanded what you could do.

Best Practices for Working With Agentic Coding Tools

Ballinger: One of nan things I’m seeing successful nan manufacture correct now is there’s a large variance successful group knowing what nan champion practices are, and moreover possibly what nan champion practices should be. That’s thing I deliberation we person to do much acquisition on, to thief group understand really to usage AI for coding assistance.

Because if you conscionable single-shot stuff, you’re not going to really beryllium successful. You person to dainty nan AI benignant of for illustration a coworker. And if you travel champion practices from improvement successful wide — things for illustration penning a personification guideline aliases a spec, things for illustration penning a method design, creating a task scheme — conscionable those 3 things, if you do that erstwhile you interact pinch AI, which you should do anyhow connected thing complex, you really person really awesome results pinch AI.

Frederic Lardinois: I was reasoning astir this: it’s still each very verbose, and we spot akin projects correct wherever we get agents.md files. Do you deliberation that is conscionable an intermediate measurement correct now successful a different direction, wherever we don’t person to do that anymore?

Ballinger: I’m not judge of that. I dream not. I deliberation I for illustration creating things. I for illustration being a craftsman and creating things, and I want to beryllium progressive pinch that. I don’t needfully want to constitute each statement of codification and debug each trial nonaccomplishment aliases whatnot, but I for illustration nan enactment of creation.

And for me, nan enactment of creation is reasoning of a problem I want to lick and past reasoning astir really I want to lick it. I deliberation AI could astir apt do that arsenic well, but you still person nan problem of, “How do I arsenic nan AI, aliases really do I arsenic a human, springiness you capable of what I want without nan AI?”

You person to do that somehow. It can’t publication your mind, astatine slightest not yet. And if it doesn’t publication your mind, past you can’t get that. And I want to beryllium progressive successful that part.

On Keeping nan Agent connected Track

Williams: How do you past make judge that it’s moving correctly? Often, erstwhile you put successful a punctual and past nan output that you get from nan prompt, and past you use it to your dev environment, it doesn’t really activity very well. So really are you verifying everything?

Ballinger: The measurement I do that is I usage nan AI to thief maine pinch that. So I don’t look astatine each statement of code. With thing for illustration Aether [Editors note: Aether is simply a programming connection for LLMs Ballinger created arsenic an experiment.], I created that architecture doc.

I did collaborate a batch pinch it. The task scheme was very minute, and aft each task, I fundamentally had it do for illustration test-driven improvement — it was for illustration TDD. It would constitute tests, and past it would instrumentality nan characteristic until nan trial passed.

I did walk clip looking astatine that. I would look a batch astatine nan tests, but I besides spent astir apt astir of my clip looking astatine nan examples. I had to create a bunch of examples covering each kinds of scenarios — record I/O, HTTP, elemental representation guidance things — and those examples, I made judge ever ran.

So aft immoderate task that was done, I would make judge nan examples ran, each nan tests passed, and I would make judge nan AI was doing those things too. In immoderate sense, it’s a achromatic container to me.

All I really attraction astir is, successful nan end, did it work? At times, I’d spell casually inspect things arsenic well.

The CLI Resurgence

Lardinois: Why do you deliberation nan CLI has go truthful celebrated again lately? Pretty overmuch each of these agents person immoderate benignant of CLI type astatine this point. Why is that?

Ballinger: I think, honestly, we ne'er abandoned nan terminal. I deliberation that’s portion of it. We didn’t deliberation astir nan terminal arsenic a superior interface. More and more, group would spell to nan terminal to tally things for illustration npm aliases conscionable different commands for illustration ls aliases whatever. But we didn’t deliberation astir it arsenic our superior interface.

More and more, location utilized to beryllium a rich | ecosystem of TUIs—terminal personification interfaces—like Midnight Commander, right?

Williams: I’m aged capable to remember.

Ballinger: Yeah, right. And it’s absorbing — we conscionable benignant of stopped reasoning astir those and conscionable had command-line scripts. But we ever usage nan terminal, and particularly successful modern development, particularly wherever there’s truthful overmuch [open root software], group usage nan terminal for git.

So it was really earthy to usage nan terminal arsenic a measurement to do these things. It gives you a batch of control. It focuses connected nan occupation to beryllium done versus tons of UI. It focuses you successful connected what you request to do. And it conscionable benignant of follows on pinch what you’re already doing aliases what you person been doing.

Williams: The CLI is an absorbing taxable now successful nan world of AI, arsenic Frederic is noting. I’m funny astir really you spot nan terminal itself evolving. I mean, nan Gemini CLI brings AI straight into it, doesn’t it?

Ballinger: Yeah, I deliberation it’s a really absorbing question: Will nan terminal germinate to beryllium much AI-capable itself, aliases will location conscionable beryllium much and much AI-capable aliases influenced bid devices wrong it?

A twelvemonth ago, I built a prototype of kubectl. I built an hold for it called kubectl AI. The squad rewrote it — my codification was atrocious — but it’s now portion of kubectl core. You tin springiness it a earthy connection query, and past kubectl tin do things pinch it, for illustration “what nodes are failing correct now,” aliases whatever. That’s 1 spot we could spot things happening—more and much devices do that.

The different point we could spot is terminals becoming much AI and agentic themselves. Warp is an illustration of that. And past also, nan astir communal shells today, absent nan terminal, are BASH and zsh. One of nan things I’ve been reasoning a batch about—I did a prototype a while back, and my team’s done different prototype that was moreover better—which is, could we create an AI-aware shell?

Each of these are ways of interacting, right? Just for illustration you could person a desktop app that understands things and you interact with, aliases a web page. Likely each of these things will happen, is my guess, because they each person absorbing usage cases.

The Future of Programming Languages

Williams: Makes sense. I’m funny again astir programming languages. Are you seeing — if we’re talking astir nan improvement of nan terminal, there’s nan improvement of tons of different things that are associated pinch nan terminal. With generative AI, you tin usage nan instrumentality to thief you make nan correct code. But is location a displacement successful really we’re reasoning astir codification itself? And pinch that, really do you spot nan improvement of caller programming languages?

Ballinger: This is simply a fascinating question, because I deliberation there’s a wide assortment of opinions connected this, and nary 1 knows for sure. I mostly person a very optimistic sentiment that’s astir apt successful nan minority. But Aether was, of course, an illustration of that.

I deliberation we should beryllium building programming languages that are designed for LLMs, wherever we ne'er look astatine them, truthful to speak. I deliberation we tin get to a spot where, arsenic a developer, I’m nary longer dealing pinch lines of code. I’m dealing pinch nan strategy design, nan architecture, nan personification interfaces, each those kinds of things. I’m crafting it together, but nan existent connection underneath is conscionable a achromatic container to me.

This, to me, feels for illustration a earthy progression of abstractions. We primitively started pinch punch cards, fundamentally instrumentality code. Then we had languages for illustration C, which were thinly connected apical of things for illustration assembler. Then we went to much and much absurd languages for illustration Java aliases C#, which continued up nan abstraction level.

The adjacent abstraction level is fundamentally my intent, what I want, and really that composes together to make an experience, to make an application. In those cases, I don’t deliberation you’ll needfully request to person nan code.

Now, I ever deliberation there’ll beryllium group whose trade is nan code, who emotion penning codification statement by line. And honestly, they’ll astir apt person a competitory advantage for definite areas of codification penning for a agelong time. That’s what they enjoy. I want them to bask it. But I deliberation we’re going to get to a spot wherever we don’t spot that arsenic much. Again, arguable opinion, I’m sure.

There was a insubstantial I publication not excessively agelong agone astir personification who put 2 agents successful connection pinch each other, and complete time, they started sending gibberish to each different because they negotiated a highly compact connection that they invented connected nan alert to talk. That benignant of point is fascinating to me. I’d emotion to spot wherever that goes.

Lardinois: Is that abstraction layer, by default, earthy connection though, aliases is it benignant of a meta-programming language? Because you want immoderate grade of precision, right? Which earthy connection isn’t — astatine slightest English isn’t ever bully at.

Ballinger: Yeah, I mean, this is simply a mobility I yet don’t cognize nan reply to. I utilized to beryllium benignant of anti-natural language. I felt for illustration location were going to beryllium different metaphors for interacting pinch AI that were much precise than earthy language.

But arsenic I’ve worked successful this section longer and longer, I benignant of consciousness like, erstwhile you look astatine a textbook aliases immoderate nonfiction novel, we don’t interest astir nan precision of what’s been written. No 1 ever says, “Well, that textbook connected calculus was imprecise.” We consciousness for illustration it was beautiful precise. And of course, there’s bad penning that’s vague, but I deliberation there’s truthful galore examples of group successful earthy connection penning precision and documentation.

I deliberation you person to benignant of meet group wherever they are, right? People could study immoderate different syntax, thing that is formal, that allows them to talk to an AI, aliases thing that’s visual, that allows them to pass pinch an AI. But past we’re asking group to study thing caller and do thing new. I deliberation you person much and much grounds that — [people should] conscionable person really bully penning skills and person nan expertise to decompose problems good and understand designs, right?

Design patterns person existed for a agelong clip now, and they’ve ever turned retired to beryllium incredibly useful, and not conscionable successful coding. Architects do creation patterns. I deliberation that’s benignant of wherever we want to go—let’s conscionable thief group beryllium amended and amended writers and amended and amended problem solvers, which, again, it’s astir decomposition of nan problem.

Lardinois: Yeah, a different benignant of skill, though, from sometimes nan accomplishment that programmers person today.

Ballinger: I would argue, I don’t deliberation that’s true. I deliberation that a awesome programmer is simply a awesome writer. They constitute down what they’re going to do. They archive nan problem, nan solution, options for nan solutions, nan design, each of that stuff. I deliberation awesome programmers do that.

And they surely understand — they’re capable to spell from business problem to method solution really well. Understanding what requirements are, knowing what’s possible, decomposing it into mini pieces truthful that they tin iterate and pivot arsenic they request to. I deliberation those are nan nonstop skills needed successful this agentic coding world. And maybe, aliases possibly not, nan expertise to constitute lines of codification is nan 1 that is not arsenic needed successful nan future.

Group Created pinch Sketch.