Bacterial Programming Meets Context Engineering: Insights for AI Agents

Andrej Karpathy introduced the term “vibe coding” and later coined “bacterial programming,” a metaphor that encourages developers to write code that is small, modular, and self‑contained, mirroring the efficiency of bacterial genomes.

1. The Three Rules of Bacterial Programming

Small : Code should be minimal to avoid unnecessary bloat, just as bacteria keep their DNA compact to conserve energy.

Modular : Functions and classes should be organized into plug‑and‑play units (operons), allowing easy composition or replacement.

Self‑contained : Code fragments should be copy‑and‑paste ready, akin to horizontal gene transfer, enabling rapid adaptation without needing the full context.

Can you imagine a function or class that anyone could “yoink” without importing new dependencies and immediately benefit from it?

If the answer is yes, you have captured the essence of bacterial programming, which promotes rapid prototyping and community‑driven evolution.

2. From Bacteria to Eukaryotes: Balancing

While bacterial‑style code excels at agility, it struggles with building complex systems. In contrast, eukaryotic genomes resemble monorepos—large, highly coupled, and organized for coordinated operation. Karpathy suggests combining the two: use a monorepo skeleton but maximize the proportion of “bacterial DNA” within it.

3. The Dependency Nightmare

Modern package managers (npm, pip) aim for modularity but often create tangled dependency graphs, turning codebases into fragile “dependency webs.” The low cost of code encourages unchecked growth, unlike the natural regularization in biology.

4. What Is Context Engineering?

Context engineering is the discipline of constructing the full set of information—prompts, tools, memories, retrieval results—that a large language model (LLM) sees before generating a response. It treats the context as a system rather than a single static prompt.

System Prompt : Defines overall model behavior, often including examples and rules.

User Prompt : The immediate task or question from the user.

Short‑term Memory : Recent conversation history providing immediate background.

Long‑term Memory : Persistent knowledge such as user preferences or project summaries.

Retrieval (RAG) : External documents or API results fetched to enrich the answer.

Tools : Functions or services the model can invoke (e.g., calendar lookup, email sending).

Structured Output : Explicit format specifications like JSON.

5. Strategies for Context Engineering

4.1 Write Context

Persist useful information outside the immediate window (e.g., scratchpads) so the agent can retrieve it later.

4.2 Select Context

Bring only the most relevant pieces of memory or scratchpad content into the window for the current task.

4.3 Compress Context

Summarize or prune the context to keep token usage low while retaining essential details.

4.4 Isolate Context

Split work among multiple agents or sandboxed environments so each handles a focused sub‑task with its own context.