My day-to-day work with AI runs across Claude, ChatGPT and Gemini. I use these tools to think through ideas, develop software and explore how to make useful work repeatable. Over the past year, I've also been exploring what happens when several agents can work on different things at the same time.
That includes hands-on work with Claude Code and Codex, experiments with Paperclip, Grok Bot and Dot, and the earlier OpenClaw-based MarkOS. These overlap. The working setup keeps changing as I try different ways to coordinate tasks.
There is a separate note on the Claude work. This one is about the coordination problem that cuts across the tools.
What I've been trying
The work ranges from gathering information and organising files to drafting documents and building websites. Some tasks can move independently. Others need an answer from me, a decision or the result of another task.
By September I was already moving between five or six AI threads, checking progress and giving the next instruction. More work could happen in parallel, but I was still doing much of the coordination: remembering which thread knew what, what had finished and what needed attention.
A Claude Code session may be implementing a change while another thread helps me think through the brief or review the result. The handoff needs to carry the decisions, constraints and evidence, rather than assuming either tool knows the other conversation.
That is the part I'm interested in improving. Dividing work between agents is easy to describe. Making the handoffs useful takes more care.
What I'm learning
A task needs a clear outcome, enough context to get started and a sensible point at which to come back to a person. “Look into this” can produce a very long answer. “Compare these options against these constraints, show your sources and identify the decision I need to make” gives me something easier to use.
The return journey matters too. I want to know what changed, what was checked and what remains unresolved. An agent's account of completing a task is a starting point for review; the resulting file, change or action is the evidence.
I still make the decisions and take responsibility for the commitments. The useful shift is how much preparation and implementation can happen around those moments.
What I haven't solved
I don't yet have a settled way to coordinate all of this. Switching between tools, keeping context current and deciding what to trust still take attention. A collection of assistants can quite easily become another collection of things to manage.
The next test is whether better handoffs reduce that effort. Can I leave a task, return later and understand its state without reconstructing the conversation? That would be a more useful measure of progress than simply having more agents running.