• Giving everyone a LLM no more replaces engineers, than giving everyone a saw replaces carpenters.

• The AI-native developer is not the engineer who ships the most. It is the engineer whose system still tells the truth about itself after the model has been working on it for six months.

• The danger is not that agents are wrong. The danger is that they are right often enough, along the visible spikes, that we stop verifying the gaps.

• A specification is a higher level of ambiguity compared to code, not a higher level of abstraction!

• Today’s bottlenecks:

  1. deciding what to build
  2. specifying what to build
  3. verifying and being accountable for what is delivered
  4. the necessary human understanding to carry out the previous three

• We stop participating meaningful in the progress of a subject when we stop being able to think fluently about it.

• Don’t be a meat proxy!

Agent’s Playbook:

• Specs constrain intent.
• Plans constrain approach.
• Tasks constrain sequencing.
• Tests, schemas, linter constrain behavior.
• Harnesses constrain execution.

My Tools:

bkmr

Knowledge Management for Humans and Agents.

twagent

Define your agents configuration canonically in TOML. Compile agent-specific artifacts for Claude Code, Copilot CLI, IntelliJ, and pi. No duplication. Handles instructions, skills, subagents and MCP configurations.

twmux

Race-condition-safe tmux wrapper giving LLM agents reliable shell I/O, a consistent JSON contract, and socket isolation.

twsrt

Translates one canonical set of security rules into agent-specific permission configs for Claude Code, Copilot CLI, and others — single source of truth, no drift.

mmdlint

Mermaid linter, ideal for agent usage in sandbox where mmdc does not work.

pi-extensions

Adds sandboxing, access control, and vim editing to the pi coding agent.