Forever Sharpening

Jordan Munch O'Hare
Self-Deprecated

Reframing Agentic Development

“Give me six hours to chop down a tree and I will spend the first four sharpening the axe.”

Myself

  • Nerd of nerds.

Myself

  • Nerd of nerds

  • Developer for developers

  • Senior Agentic Developer (SAD) at Self-Deprecated

SAD times

  • “Programming is disappearing”
  • “How can people learn to code today?”
  • “I miss real development”
  • "Too much context switching. Easy to lose track."

Past changes

  • Punch card programming

Past changes

  • Punch cards gave way to screens
  • The work moved up an abstraction
  • The programmers did not disappear.

Another interface shift

  • Agents perform more of the work
  • Agents provide first-pass judgment
  • We remain accountable.

This time, we can shape the tools

  • Output
  • Tooling
  • Capacity to learn.

Move up in agency

Wait for the next release
Work around issues
Experiment
Build your own
Incorporate & iterate
PASSIVEACTIVE →
  • What does that look like?

1. Work concurrently

  • Many sessions/tabs/windows

1. Work concurrently

  • Many sessions/tabs/windows
  • Whatever works for you: Niri workspaces · terminal tabs · Claude Cowork
  • Juggling

2. Make the loop self-healing

  • Problem: it comes back every five minutes
  • Change: make checks deterministic
  • Loop: edit → test → repair → retry
  • Web work: see error → fix → rebuild → inspect
  • If it must happen, do not ask the model to remember it.

3. Choose the kind of work

Pick your battles:

  • Exploratory work
  • Main work / deliverables.

Explore

MVPs · POCs · tracer bullets

  • Fast, flexible, disposable
  • Less focus. Pick up when you have time
  • Use skills or extensions like Ponytail.

Deliver

Important work with an expected outcome

  • Grill-me / Wayfinder for in-depth planning beforehand
  • Full self-healing
  • Main ball in the air.

4. Manage your attention

Which agent needs me?

  • Don't keep deciding, or have to find them. Make it part of the loop
  • pi-attention
  • Context switching, but the good kind.

5. Reduce the cost of judgment

Wall of text output

  • Pre-planning can reduce this
  • Get the output in HTML, with the rest collapsed
  • Make acting on the result as cheap as possible
  • i-have-adhd skill.

Stay close to the code

Where appropriate

  • Hunk: interact with the agent’s changes

Stay close to the code

Where appropriate

  • Hunk: interact with the agent’s changes

  • Stacked diffs: review logical chunks, upstream and locally.

6. Manage context

  • Larger context windows give worse results
  • Agents forget instructions in the middle more
  • More context = more usage.

Control what enters

  • Persist lessons into AGENTS.md
  • Rewind: Double Escape
  • Delegate bounded work: subagents
  • Compact when history stops helping
  • Filter output: RTK returns actionable failures.

Working across many repos

  • Monorepos: relationships are visible
  • Hubs: coordinate independent repositories.

7. Manage usage

  • Plan with stronger models
  • Implement with cheaper models
  • Review cheaply first
  • Challenge it with Codex adversarial review
  • Escalate judgment
  • Spend expiring capacity deliberately.

Use it or lose it

“I have five hours left, and 30% on my subscription.”

  • Turn on fast mode
  • Start a swarm
  • Update documentation
  • Do discovery work
  • Plan the next big task
  • Improve your process.

User as a multiplier

Your skill × agent capability = agentic development output
  • Judgement is our main skill
  • Learn day by day, change by change
  • Sharpen your own knowledge, don't become dull.

Move up in agency

Wait for the next release
Work around issues
Experiment
Build your own
Incorporate & iterate
FIND YOUR FRICTIONMOVE ONE STEP →

Forever Sharpening

“Give me six hours to chop down a tree and I will spend the first four sharpening the axe.”

Often attributed to Abraham Lincoln

  • No need to choose. Do both at once
  • Sharpening is the means, not the end.

Takeaways

  • You’re in the driver’s seat
  • Agents multiply the skill you bring
  • Take agency in your own development.

Not obvious?

  • Ask your agent
  • Ask me
  • Find the links online.

Self-Deprecated

  • 1:1 sessions
  • Team workshops
  • Next month’s workshop
QR code for selfdeprecated.ai/hello
selfdeprecated.ai/hello

Talk: Claude Code for Developers, Sønderborg. 15 minutes + 5 minutes for questions.

Establish technical credibility, then show each tool full-screen before returning here.

My Moonlander keyboard.

Return to the original slide and finish the introduction.

Acknowledge the anxiety rather than dismissing it.

Introduce the historical interface, then show the old and current interfaces full-screen.

Public-domain Argonne/US Department of Energy photograph via Wikimedia Commons: HD.6B.030 (10693137083).

Return to the argument: programming was redefined rather than erased.

Agentic development changes where the developer applies skill; it does not remove responsibility.

The distinction is ownership, not which feature someone uses. Using a skill is different from identifying a need and creating the skill you wish existed.

This is not binary. Move one step when recurring friction earns the investment.

Set up the problem, show familiar Claude Code concurrency examples, then demonstrate my Niri workflow.

Boris Cherny describing five parallel Claude sessions with numbered tabs and notifications.

Full-screen Niri demonstration. The poster is the PDF and venue fallback.

Return to the slide and explain that the interface is incidental.

Run tests through hooks and return failures to the conversation. For web development, give the agent browser access so it can observe the UI or console error, fix it, run the dev build, and inspect the result itself. Longer loops make human attention scarce.

Explain normal notifications versus an ultra-important completion immediately focusing its workspace. Once results arrive reliably, reviewing them becomes the bottleneck.

Show Markdown wall versus a structured HTML report. Do not accept the default interface between agent output and human judgment.

Hunk split view with sidebar and inline AI notes. Source: modem-dev/hunk README.

Hunk stacked display mode and mouse-selectable menus. Source: modem-dev/hunk README.

The second screenshot is Hunk's stacked display mode, not a logical stacked-diff workflow. GitHub stacked-diff support is currently beta.

RTK token-savings report.

Shared scope does not mean loading every repository into model context. Let the agent discover only what the task needs.

Mention GLM/Grok first-pass review verbally, then a stronger model for the final decision.

Expiring usage can sharpen tomorrow's work without distracting the primary task.

Nobody needs my exact setup. Everyone can make one deliberate move towards ownership.

Agentic work lets one process cut while another sharpens.

Tesla may drive to A, but it cannot decide to go there or judge whether it was correct. Keep the user's preferred force-multiplier/enumerator phrasing.