Leaning In to Find Out

Who Am I?

• Armin Ronacher, co-founder at Earendil
• Spent 10 years making Sentry a great product
• Created a lot of Open Source software: Flask, Jinja2, and others
• I write at lucumr.pocoo.org about agentic engineering

We Are All Somewhat Beholden to Foundation Model Providers.

That Means We Have to Learn From Behavior, Not From Internals.

The Weird Part

• As user we do not control the training process
• We rarely know what the models were optimized for
• What if the models change and our uses perform badly?

Basics

Understand Reinforcement Learning

• A pretrained model gives you a broad prior
• RL then samples behavior and attaches a reward signal
• Good trajectories get reinforced, bad ones get suppressed
• Over time, the model becomes more like the behavior that keeps scoring well

What Do the Model Providers Train On?

In short: entire sessions they get access to.

• prompts, responses, tool calls, edits, retries, tests
• ranking signals: task success, acceptance, abandonment
• particularly things like Claude Code sessions where the loop is rich
• As we vibe, we reinforce

So How Do We Extrapolate?

• Do not just ask what is popular
• Ask where behavior is abundant and legible
• Ask where outcomes can be scored cheaply
• Ask which environments dominate those traces: shells, files, test loops

Bonus Hunches

• Coding agents increasingly bring screenshots into the loop
• That means visual understanding in the context of code or code driven agent turns
• And coding-adjacent tooling is becoming the base layer for workspace automation

Layers Above the Coding Substrate

• The same agents now manipulate spreadsheets, Word documents, calendars, and similar workspace surfaces
• They are also remote controlling tmux sessions and web browsers via CDP
• So the coding-agent substrate is starting to absorb more and more of ordinary computer work

Surprises

• Curious failure modes show up when important state changes are not visible in the session transcript
• Agents often get very confused when state changes outside the turn and the harness does not inject those changes
• If you tell them about a change and ask them to validate it, they often hallucinate and claim they called a tool when they did not

What This Means for Us

• Do not just optimize for what the model can do today
• Build around workflows with dense textual feedback
• Expect the substrate of coding to keep improving: writing code, running it, files
• That gives you a strategy that survives model drift for a while

Coding Is a Mode of Execution

• If the file system works, and agents know how to program, you can route a surprising amount of work through that loop
• Many "non-coding" tasks become tractable once they look like files, programs, tools, and outputs
• That gives you leverage from the model's strongest priors instead of fighting them

Build Tiny Programmable Worlds

• Use simple sandboxed languages with basic sandboxes
• Prefer text files, explicit inputs, deterministic outputs, and good error messages
• If the agent can iterate by editing, running, and reading failures, you inherit the coding-agent gradient

Alignment Beats Novelty

• SQL beats a custom query DSL
• JavaScript light beats a complicated expression language
• Even a bespoke DSL can work if it aligns with a known language and produces excellent errors
• The more it feels like programming, the more capability you get for free

Initial Pass, Then Iteration

• Ask the model a few times how it would do the thing
• Models have some weak introspectability about likely strategies and failure modes
• Then feed agentic turns back into models so they can critique the interaction and suggest harness, prompt, and tool improvements

One Practical Loop

• Run the same basic task twenty times with a coding agent
• Use subagents or small variations if that helps you explore the space
• Then let another LLM score the outcomes for quality, failure modes, and consistency

Use Your Own Traces

• Your own coding sessions are a valuable input for judgement
• Things like .{pi,claude,codex}/sessions accumulate a surprising amount of useful data
• At the beginning, reduce variables: fewer tools, simpler harnesses, clearer prompts