Summit notes: This is Good For Humans!

One of my favorite books is Tracy Kidder’s “Mountains Beyond Mountains,” chronicling physician Paul Farmer’s challenges providing care in rural Haiti. There is a proverb in that book that is the right framing here:

Beyond mountains, there are mountains.

I started with a simple goal, iterated, and found one route up Agentic Mountain (working title; not new: naming is hard).

At the top there is a feeling of accomplishment and relief. There is also one of both humility and exhilaration looking at what’s ahead, and of curiosity to learn the routes that others will undoubtedly discover.

None of this should be read as prescriptive or gospel; it’s just one route I found. And I sincerely hope and trust that this is a journey we embark on collaboratively, together. This is a group climb, not an arms race.

Er, we’re climbing, so our arms race. Anyway. Enough Dad. Let’s climb.

Base Camp

All routes up this mountain start with a single axiom:

I don’t write code anymore.

It’s not enough and it’s all you need.

It’s everything old is new again.

You’ll see what I mean.

Excelsior.

Beta

• Some double entendres write themselves. In climbing, beta is crucial info about routes, difficult sections, good holds, optimal moves.
• Please don’t ever ask me to actually climb with you---I just like wordplay and am terrified of heights.
• I’m intentionally keeping this tool-agnostic.

• Choreography: Crafting the prompt.
  • Expand on Requirements: The more specifics, the better.
    • But: Sometimes an agent will hallucinate, AND sometimes will come back with a novel solution. Always follow a line of inquiry just like you would with a junior dev. I’ve had agents not follow the solution I’d proposed. I pulled that thread and realized that the agent was, in fact, correct. There was an edge case I hadn’t accounted for. Be open to that!
  • Testing as Forethought: Probably 50% of my initial prompt was detailing test cases. Testing was integral to the plan rather than the afterthought it often is.
    • Not new: This is how it SHOULD be!
  • Reference Repos: Something I didn’t think to do is to point the agent toward a different repo that had already solved a similar problem. i.e. Gradle upgrades, major module version bumps, etc.
  • Paraphrase: Ask your agent to paraphrase what it’s about to do. Steer that prompt until it meets what you’re looking for.
• I to We: What does it mean if I’m a team now? How do I look at my backlog differently?
  • Divide and Conquer!
    • Human teams work in parallel.
      • Not new: Brook’s Law---throwing more agents at the same repo doesn’t necessarily make things move faster.
        • There may be a different route up this mountain where agents can sort out the merge conflicts, but I wanted to start with the low-hanging fruit summits.
      • I grabbed tasks from three different projects at a time.
      • Projects that required serial work had their agents run serially. The next agent would run after the previous agent got its work into PR, on a branch off of the previous agent’s branch. They were instructed to indicate that the previous PR was a blocker.
  • Watch (cognitive) overhead:
    • A central lesson in the John Wick films is that “everything’s got a price.” Another lesson is that what appears to be a hero story is, in fact, a Greek Tragedy. Every Greek Tragedy has a katabasis---literally, “descent.” - Not new: Don’t be a hero. - Context-switching has a price. You’re still human. Respect your limits.
    • I found that my sweet spot was running three agents in parallel across different repos. Two resulted in idle time; four was too much.
      • Since agents take about 20 minutes to return results, I rotated my attention in 20-minute cycles. There are some neurological optimals here. That’s about what the science says that it takes to get ramped back up on a task. The benefit here is that the agent is tracking the context for you, so the switch isn’t as taxing. But it’s still there.
    • There’s no wrong answer! You’ll discover your optimal. The biggest benefit comes from using agents in the first place. It’s more important to not overload yourself.
      • Not new: Amdahl’s Law---parallelism scales logarithmically.
    • This is active. I needed to review every change. I found places to DRY stuff up, saw refactors, found variable names that didn’t quite fit. I treated each iteration like a PR and re-prompted the agent with all my feedback.
      • Not new: This is still Deep Work; full engagement; flow state.
• The Verification Loop: I had the agent run the full integration suite after every change. If it broke something, it had to fix it before coming back to me.
  • Self-Correction: The beauty of this is that if the tests fail, the agent sees the error and iterates. I found that by focusing the bulk of my prompt on testing requirements, the agent returned results that were actually verified, rather than just “looking right.”
• End-to-End Orchestration: I had the agents handle everything from creating the branch to pushing the PR. I never touched code or a terminal between those points.
• SUPER SUPER SUPER PRO BETA POWER MOVE: Your agents can create MULTIPLE PRs.
  • Take a second.
  • Think about that.
  • It’s powerful. We know (and studies confirm) that bugs get lost in long PRs.
  • Humans get overwhelmed. We have our own context window. The industry has tried and failed to address this.

The Summit

What I see behind me.

1. THIS IS GOOD FOR HUMANS!
   1. These developments are the ultimate ascension of the core principles and best practices we’ve held dearly as a craft, but haven’t been fully empowered to live up to until now.
   2. Testing is no longer an afterthought. It takes on central importance and is a part of the prompt.
   3. WE ARE WRITING CODE FOR HUMAN CONSUMPTION. ALWAYS HAVE, ALWAYS WILL.
      1. LLMs are “human-imitators.” Even when you’re writing for an AI, you’re writing for something trained on human logic. Clear variable names and documentation aren’t just “nice to have” for us---they are the cairns (guideposts) for the agent.
2. Our processes are often broken. They’re designed for individual humans, not for agentic teams. In the immediate term, this means addressing process bottlenecks. In the longer term, this might mean a fundamental rethinking of our roles.
3. Gravity is different up here. We’ve gained weight. Engineering teams have traditionally run lean. In the immediate term, we have increased bandwidth to reduce risk, increase quality, ship better products. In the longer term: we might go lean again to increase the scope of what we deliver.

The Mountains Ahead

• Building out agent skills: combining prompts with deterministic scripts to handle things like fetching module versions, fetching tokens, enforcing code styles, etc. We can DRY out OUR TIME. Focus on the important stuff.
• The first bottleneck: Code review.
  • Even if I’m coding at the speed of an agentic team, that code still gets reviewed at the speed of a human. Not new: we still need human code reviews.
  • But what if we could offload some cognitive overhead there? Have agents be our guardrails?
    • Agents that remind us about our naming conventions, catch misspellings, read docs and call out risks we didn’t realize (actual thing a friend ran into this week when an agent caught him making a library call that had performance implications)?
    • What if we could focus on the readability, the logic, etc? Again, this is a positive development! This is what we’ve aspired to! This is good for humans!
• The agents need more context. What if we add MCP as a requirement for each project?

The Horizon

• What is our new value proposition?
  • Well, who is on my team?
  • LLMs are good at language. What does that unlock? Everything.
  • What if my team of agents isn’t just coding agents in my own domain?
    • What if it’s full-stack, UX, QA? Mobile? Polyglot?
    • Does that lead to a new value prop? More specialized products? Think Different Agentic.
    • Is there unmet need for more product lines?

There is much more to explore. This was just the first ascent, and the oxygen up here is not thin but rich. The territory ahead full of promise.

One more book reference.

Michelle Obama’s “Becoming.”

Let’s frame Ascending in the same way here. Ascending is the state of being that we’re in. Perpetually.

Not new: What first drew me to this field was the pure, human joy in building, learning, growing, and solving problems.

Agentic tools are not anathema to that; rather, they are in complete harmony with it.

We are Ascending. Always have been.

Unleash the Humans.

Send it!