Agents that narrate their work are the best team players – Jon Udell

7 hours ago

Agents like Claude Codex and Codex run in the terminal, wielding system tools (awk, bash, curl, git, python) and MCP tools.
Bram's latest feature uses Haiku to convert commands, tool calls, filenames, and line numbers into statements of intent, making agent transcripts more readable.
Narrating work—whether by humans or agents—is essential for transparency, teamwork, and learning, as seen in open source development and internal blogging.
AI agents are uniquely qualified to narrate their own work, and proper harnesses can elicit this behavior, transforming how we observe and understand agent actions.
The approach builds on earlier methods (e.g., using small models for categorization) and is presented as a transformative improvement for agent-assisted workflows.

Knock Agent is an AI agent for managing customer messaging resources, capable of creating and updating workflows, templates, audiences, segmenting users, and answering performance questions.
The agent was built using a virtual filesystem and bash, inspired by Vercel's approach, allowing it to explore and script on resources via a filesystem instead of exposing numerous API tools.
A custom Elixir-based virtual bash interpreter called just-bash was developed to provide a lightweight, in-memory environment without the performance overhead of a full sandbox.

Adopting a new tool involves three phases: inefficiency, adequacy, and life-altering discovery.
Chatbots are limited for coding; use agents that can read files, execute programs, and make HTTP requests.
Reproduce manual work with agents to build expertise, breaking tasks into clear, actionable steps.

For many agent tasks, the best 'tool' is a small code-execution surface, allowing models to write ordinary code against typed capabilities and enforce security below the model.
Traditional agent toolboxes are growing with many tools, but models already excel at code, which is a more familiar and efficient interface, reducing context usage and tool calls.
Code mode compresses context (e.g., 88% less), reduces token consumption from tool schemas, and allows deterministic logic between calls, such as joins, filters, and error handling.

The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.

The author now relies heavily on AI agents for producing entire PRs in familiar areas, starting every change with an agent request and pushing the PR after a single editing pass.
Agents are used extensively for bug investigation, successfully diagnosing 80% of issues autonomously, though human expertise remains crucial for narrowing down complex bugs.
The author still writes PR descriptions by hand to signal human review and avoids using AI for Slack messages, ADRs, or blog posts, but uses AI for feedback on drafts.

Knock Agent is an AI agent for managing customer messaging resources, capable of creating and updating workflows, templates, audiences, segmenting users, and answering performance questions.
The agent was built using a virtual filesystem and bash, inspired by Vercel's approach, allowing it to explore and script on resources via a filesystem instead of exposing numerous API tools.
A custom Elixir-based virtual bash interpreter called just-bash was developed to provide a lightweight, in-memory environment without the performance overhead of a full sandbox.

Adopting a new tool involves three phases: inefficiency, adequacy, and life-altering discovery.
Chatbots are limited for coding; use agents that can read files, execute programs, and make HTTP requests.
Reproduce manual work with agents to build expertise, breaking tasks into clear, actionable steps.

For many agent tasks, the best 'tool' is a small code-execution surface, allowing models to write ordinary code against typed capabilities and enforce security below the model.
Traditional agent toolboxes are growing with many tools, but models already excel at code, which is a more familiar and efficient interface, reducing context usage and tool calls.
Code mode compresses context (e.g., 88% less), reduces token consumption from tool schemas, and allows deterministic logic between calls, such as joins, filters, and error handling.

The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.

The author now relies heavily on AI agents for producing entire PRs in familiar areas, starting every change with an agent request and pushing the PR after a single editing pass.
Agents are used extensively for bug investigation, successfully diagnosing 80% of issues autonomously, though human expertise remains crucial for narrowing down complex bugs.
The author still writes PR descriptions by hand to signal human review and avoids using AI for Slack messages, ADRs, or blog posts, but uses AI for feedback on drafts.

Hasty Briefsbeta