I currently divide agent work between two assistants: Grok Bot for personal systems, and Firstmate for software development. This note records how each is used, the principles I follow when managing an agent team, and where the approach still falls short.
Personal systems: Grok Bot
Grok Bot handles personal systems and operations: administering a remote VPS, aggregating information from several social platforms, maintaining this site, and handling campus affairs.
Development: Firstmate
For software development, I use Firstmate by Kun Chen (kunchenguid) to manage the agent team. His YouTube channel, @kunchenguid, covers the workflow in more detail; “High Throughput Agentic Engineering with Kun” is a good starting point.
Managing an agent team
Chat logs and markdown alone are a weak basis for reviewing work. I ask assistants to report with presentable material, for example through Kun’s Lavish. Two further rules apply:
- Use the tool. When a tool is available, the agent should actually use it rather than describe what it would do.
- Bring evidence. When a tool fails, the agent should return concrete evidence rather than a bare report of failure.
In frontend work, this means generating page samples with GPT’s image mode before implementation, and using Chrome screenshots during the build as the basis for adjustments.
Kun Chen’s post on the executive mindset takes the same position from the standpoint of running an organization. Its main points, paraphrased:
- Once a task is assigned, there is no need to keep watching the conversation.
- Important decisions should rest on documents or presentations, not on chat alone.
- When something goes wrong, find the root cause rather than correcting only that one case.
- A manager should not become the testing bottleneck.
- Before deciding whether to step in, ask: if I had 100 times more agents doing 100 times more work every day, how would I handle this?
— Paraphrased from Kun Chen, “executive mindset,” X, 13 November 2026 (Asia/Taipei)
Research
In research, my use of agent teams is still exploratory. I cannot yet fully trust agents, so the work still requires substantial personal involvement.