My last post captured my feelings and thoughts on AI. It’s also worth capturing how I use AI, specifically GenAI (as if that needs saying at this moment).
If I were to place myself on the Steve Yegge Gas Town 8 Point Scale of AI programming, I’d place myself between 5 and 6. I run Claude Code in the background and regularly give Claude deep research assignments.
At work, IBM has only just granted me access to its coding agent, IBM Bob. But I’ve been able to start quickly by using my experience using Claude Code at home.
Let’s list out a few specifics:
Agents are clearly where it’s at, but not completely
By far my biggest interaction surface with AI is coding agents, Claude Code and IBM Bob. But I still end up using Claude’s chatbot interface for questions and for kicking off deeper research. Even though a coding agent’s interface is focused around chat, the hyper-optimised chatbot UX still wins out for quick questions — perhaps it’s just that it loads faster.
I don’t use AI IDEs
I’m a terminal user. A Helix user. Left pane editor, right pane Claude Code works well for me, as does having background terminal tabs with other Claude Code instances working. Thankfully IBM Bob has a shell UX, though it feels like the IDE gets most of the love.
I think that having the agent slightly less integrated than in something like Cursor helps me. The agent has its own place. That helps me think more clearly. This place for reading and writing code, that place for talking.
I think this helps me think of the agent as something that can do more than write code. I do a ton of stuff in the terminal, and having the agent live in that environment reminds me of its flexibility.
I don’t really use MCP servers if I can avoid it
Agents can do a lot with the command line. Yes, I need to try out Pi.
I’m not super-religious about this, though I am wary of context bloat through MCP use. I did build MCP support into rapport, though I never used it much.
I build my own sandbox for Bob
Agent security is a problem. Most agents (all agents?) don’t really seem safe.
However, without yolo mode it feels like the agent is putting you to work: “Watch this stream of confusing junk, and be continually responsive to my prompts”. To improve this situation, I used Anthropic’s sandbox-runtime to build a better sandbox for Bob. In particular, the network is locked down to decrease the chance of nasty stuff polluting the context.
I’d like to upgrade to a limactl VM, but need to find the hour or two to get it working smoothly. After that, it’d be nice to do an internal presentation on safer yolo-ing.
I am interested in using coding agents for non-coding use-cases
I find it particularly interesting to ask: what non-coding tasks can I get a coding agent to do? Coding agents are thin wrappers around hugely capable models. How can I get the agent to do stuff I find tedious?
At work I’ve built a skill for Bob that smartly processes GitHub notifications. It uses the
ghCLI tool (which the models underlying Bob are pretty good at).Working through notifications with Bob is a surprisingly nice experience. I have Bob pull the description and comments, and summarise the state of each issue in turn. I always found scrolling around each issue to figure out what was going on with it tiring. It’s great that Bob can do this for me. And I can handle everything from the keyboard in natural language — no remembering GitHub’s invisible hotkeys.
But the killer feature is the smarts of the model. You can break out of the fixed workflow:
- “Mark all the notifications from this repo as read”
- “Remember this issue is important to me”
- “Hang on, I don’t get this particular type of caching we’re talking about. Can you give me some background on it?”
- “Could you give this PR a quick review, let me know your findings?”
I want to build more of this kind of thing — “coding adjacent” tasks.
Obviously this kind of thought process is what took Anthropic to Cowork. The core agent loop combined with tools that can affect local data — augmented with MCP servers — is a powerful engine.
I fear losing my skills
Writing prose and code atrophies if not practiced. I fear losing capability if I delegate to AI. I want to retain these skills — I enjoy both — while not giving up on AI. To solve this I am doing two things:
Making time to write both prose and code without AI. Particularly when trying to learn new things. Doing something beats reading about it.
Learning to effectively build agent harnesses. I want to be able to delegate work to agents, and have them write code I can be confident in without my having to review every line. Reviewing code is hard, and reviewing the quantity of “decent looking” code that AI pumps out is nigh on impossible.
This feels like the gordian knot of AI — those that solve it will reap rewards without burning out.
Overall, I still find myself conflicted. Skill loss is real, I don’t want to spend all my time reviewing agent PRs and I don’t want to get burned out:. Thinking about how to find the balance of productivity and enjoyment at work is where I’m at right now.