The Knowledge Chipper: An Agentic Coding Story

5 hours ago

Developers and LLMs spend significant effort building context to make code changes, but most of that knowledge is lost after each session.
The article 'The Session You Cannot Take With You' argues that black-box AI agents damage institutional knowledge and session resumability.
Different LLMs like Codex and Claude working on the same code start from scratch, wasting millions of tokens.
LLM portability becomes urgent due to geopolitical disruptions, such as the AWS datacenter outage in Bahrain.
Combining LLMs (e.g., Google's agent reviewing Claude's work) highlights the inefficiency of context loss.
In larger teams, code reviewers lack context from the original LLM session, leading to poor understanding of nuanced changes.
Junior developers may produce large code changes without documenting context, forcing senior engineers to recreate all context from scratch.

LLMs can write code for you, reducing manual effort but leading to duplicated logic.
Repeated conditional checks in different files create maintenance issues.
LLMs learn from existing code patterns, reinforcing bad practices if shortcuts are used.

The author chooses not to use LLMs for coding, not even tools like Claude Code or Cursor.
The author values the craft of programming, arguing that friction in code-writing encourages thoughtful design.
Using LLMs is seen as surrendering autonomy, creating dependency on corporate tools that may increase costs.

Software engineers have differing views on codebase understanding: small, low-turnover teams advocate complete understanding, while large, high-turnover teams accept partial understanding.
Peter Naur's 'Programming as Theory Building' argues that code is a by-product of a programmer's 'theory of the program,' but his conclusion that code should be discarded when theory is lost is impractical for large systems.
Large software systems cannot be rebuilt from scratch due to countless quirks; rewrites must occur incrementally by carving out small chunks.

LLMs represent a different kind of intelligence, distinct from human intelligence.
Human analogies like 'junior engineer' or 'abstraction layer' are flawed because LLMs lack intent and deterministic reasoning.
LLMs combine vast knowledge with an inability to understand intent, leading to subtle drift from user goals over time.

AI adoption is often mandated by management without engineering background, causing resistance, but the author found it useful for search and summarization.
Hallucinations are inherent in LLMs; always verify primary sources, especially when using AI for coding or research.
LLMs are poor at writing code, especially for game development, due to limited and poor-quality training data, often producing over-engineered solutions.

LLMs can write code for you, reducing manual effort but leading to duplicated logic.
Repeated conditional checks in different files create maintenance issues.
LLMs learn from existing code patterns, reinforcing bad practices if shortcuts are used.

The author chooses not to use LLMs for coding, not even tools like Claude Code or Cursor.
The author values the craft of programming, arguing that friction in code-writing encourages thoughtful design.
Using LLMs is seen as surrendering autonomy, creating dependency on corporate tools that may increase costs.

Software engineers have differing views on codebase understanding: small, low-turnover teams advocate complete understanding, while large, high-turnover teams accept partial understanding.
Peter Naur's 'Programming as Theory Building' argues that code is a by-product of a programmer's 'theory of the program,' but his conclusion that code should be discarded when theory is lost is impractical for large systems.
Large software systems cannot be rebuilt from scratch due to countless quirks; rewrites must occur incrementally by carving out small chunks.

LLMs represent a different kind of intelligence, distinct from human intelligence.
Human analogies like 'junior engineer' or 'abstraction layer' are flawed because LLMs lack intent and deterministic reasoning.
LLMs combine vast knowledge with an inability to understand intent, leading to subtle drift from user goals over time.

AI adoption is often mandated by management without engineering background, causing resistance, but the author found it useful for search and summarization.
Hallucinations are inherent in LLMs; always verify primary sources, especially when using AI for coding or research.
LLMs are poor at writing code, especially for game development, due to limited and poor-quality training data, often producing over-engineered solutions.

Hasty Briefsbeta