When AlphaFold Meets Claude Code and Codex in the Lab Notebook

By Muhammad Kashif · Draft · AI automation & cybersecurity builder, Pakistan

AlphaFold sitting next to Claude Code and Codex in the same notebook week looked like a category error until I read it as a lab-notebook problem: multi-agent teams, hash-line edits that touch only broken lines, persistent memory, and a debugger that does not rewrite the universe to fix a comma. Science software taught provenance. Coding agents are finally being forced to learn the same manners.

Five repos, one discipline

The notes pointed at a Karpathy-style free guide on coding with AI, a free full-app-from-one-sentence builder as an alternative to expensive monthly tools, AlphaFold as a Nobel-era reminder that specialised models change fields, and add-ons like OmiPy for Claude Code and Codex. I do not use AlphaFold in client web apps. I do copy the idea that a tool should have a clear input contract and an evaluation story beyond "it looked right."

Free builders are for learning and spikes. Nobel-grade systems are a humility check: domain models beat vibes when the domain is real.

OmiPy-shaped multi-agent manners

Multi-agent teams, hash-line edits with anchors per line so fixes touch only broken lines and save tokens, a built-in debugger, persistent memory — that is the manners list I want in any coding harness. Whole-file rewrites for a one-line bug are how reviews drown and how subtle correctness dies.

I configure agents to prefer anchored edits. When a harness cannot do that, I shrink the task until it can, or I switch harnesses.

Lab notebook rules in a git repo

Every agent run records model id, skill version, command, and whether tests passed. Failures are stored, not overwritten by the next hopeful attempt. That is lab culture, not bureaucracy. Without it you cannot tell whether you improved the system or got lucky once.

AlphaFold's public story is about prediction quality under scrutiny. My coding agents deserve scrutiny too, even when the stakes are a checkout bug rather than a protein.

What I refuse to romanticise

I refuse to claim biology expertise I do not have. I refuse to use AlphaFold as marketing glitter on a CRUD app. I refuse one-sentence app builders for anything that handles money or identity without a graduation path into a real repo with tests.

Romanticism is how workshops oversell. Precision is how retainers renew.

Debugger before another agent

When something fails, I want a debugger loop, not a fourth agent offering opinions. Persistent memory should remember the failing test and the last diff, not a novel about what the agent "felt" went wrong.

OmiPy-style persistence earns a seat if it shortens mean time to green. If it only lengthens chat, it is clutter.

Token thrift as engineering ethics

Hash-line edits saving tokens are not just cheapness. They reduce accidental churn in unrelated lines — the kind of churn that creates merge pain and hides security diffs. Thrift here is a reviewability feature.

I measure churn: lines touched versus lines needed. Agents that spray lose privileges.

How I brief a lab-shaped coding pilot

Pick one service. Golden tests first. Anchored edits required. Memory scoped to the service. No production secrets. Human review on auth. Two weeks of logs before expanding to a second service.

If the pilot cannot stay inside those rails, we do not have an agent problem. We have a discipline problem, and adding AlphaFold metaphors will not fix it.

What leaves the notebook

Anchored edit preference, failure retention, model pins in PRs, and humility about domain tools. The Nobel slide stays inspiration. The manners become defaults in AGENTS.md.

Anchored edits in code review

I ask reviewers to flag whole-file agent rewrites the same way they flag missing tests. OmiPy-style hash-line discipline becomes a social norm, not only a tool setting. AlphaFold metaphors stay out of the review comments; manners stay in.

When a junior asks why we care about anchors, I show a diff where an agent "fixed" a bug and quietly rewrote a rate limiter. Once is enough teaching.

Failure museum

I keep a folder of failed agent runs with notes — including ones that looked green until a human noticed. Lab notebooks need negative results. Codex and Claude Code learn from what we store; deleting failures teaches optimism bias.