hebel. today · 2 November 202662 / 69
Science · Data Science Applications
Shift AI investment from token volume to shipping velocity
Measuring AI success by token consumption creates a misleading financial picture. While agents increase token usage by planning, testing, and iterating on tasks, spending caps often restrict productive workflows rather than curbing waste. High…
From 4 newsletters · 1 min read · one thing to try
Measuring AI success by token consumption creates a misleading financial picture. While agents increase token usage by planning, testing, and iterating on tasks, spending caps often restrict productive workflows rather than curbing waste (MIT Technology Review). High token counts frequently reflect runaway loops or excessive retries rather than actual software output. Instead of rationing usage, teams should track the cost per merged pull request, calculated as total AI-assisted development costs divided by the number of successful, reviewed changes. This metric highlights whether increased AI usage converts into finished software or merely accumulates technical debt and maintenance costs (Tech World With Milan Newsletter).
Route routine tasks to open-weight models and reserve expensive frontier models for novel reasoning. Open-weight models now approach frontier performance on specific engineering benchmarks, making them suitable for code generation, test creation, and refactoring. Establishing a gateway to route requests based on task complexity and cost prevents premium inference pricing from inflating entire workflows. This division requires evaluating models against your specific codebases and acceptance criteria before assigning production work, ensuring that infrastructure costs align with the value of the delivered output.
Effective AI integration requires treating process changes as migrations rather than ad-hoc experiments (Refactoring). When moving to agentic workflows, identify the specific problems current processes solve to avoid creating new friction. Involving team members who are most invested in existing tools helps address underlying concerns and secures necessary buy-in for new methods. Once a direction is established, platform teams should draw hard lines by standardizing on specific agentic tools, providing dedicated support for those choices while leaving others unsupported. This structured approach ensures that AI adoption improves delivery speed rather than just increasing the number of open pull requests.
Try today · 20 min
Create a repository to document agentic lessons and rules.
Start with two files: AGENTS.md for your agent's role and instructions, and context/lessons.md to record recurring corrections.
You need
- A private git repository
- AGENTS.md file
- context/lessons.md file
Steps
- Initialize a new private repository.
- Create the AGENTS.md file to define the agent's primary role and constraints.
- Create an empty context/lessons.md file.
- Commit these files to the main branch.
- The next time you correct an agent twice for the same issue, document the fix in lessons.md.