This is incredibly exciting to read! Also I'm quite envious of your token usage. :D I want to know all the details on how your team learns and fails using lobsters for their own work. The failures I think are important too. For example, my agent moved a meeting on me and denied doing it. We had an actual disagreement and I felt gaslit honestly. But 'he' updated his file and swears it won't happen again. These are important learnings and I feel like we all learn through someone else's experience! Looking forward to more updates.
Really enjoyed this one—felt like a clear signal of where “personal agents that actually do things” are heading.
I ended up writing a response riffing on a slightly different angle: the gap between having an AI and actually compounding advantage from it (what I’m calling the “Copilot Ceiling”):
I love the dev model: it is the same used by Anthropic and OpenAI and, in the fast moving AI environment, it is the only one that works, also from personal experience.
The exciting challenges, besides the security implications, are customization and token efficiency. Hardware diversity makes a fit for all solution genuinely hard, but that's also what makes the space interesting: there's real room for innovation. And while the cost question is real, prices have been falling fast, and what feels expensive today will likely look very different in 12 months.
Your earlier post about AI actively organizing a desktop — not just searching it — became the origin story of a project I recently built: a two-agent MCP system for job application lifecycle management inside Claude Desktop. One agent is human-in-the-loop for resume customization. The second watches Deleted Items in Outlook — when I delete a rejection email, it autonomously archives the matched folder and logs it.
"Everyone has a Lobster." I've been living this, not just reading about it.
Congrats on the new role — excited to see Ocean 11 take shape.
This is incredibly exciting to read! Also I'm quite envious of your token usage. :D I want to know all the details on how your team learns and fails using lobsters for their own work. The failures I think are important too. For example, my agent moved a meeting on me and denied doing it. We had an actual disagreement and I felt gaslit honestly. But 'he' updated his file and swears it won't happen again. These are important learnings and I feel like we all learn through someone else's experience! Looking forward to more updates.
I think your Lobster needs a red badge... until it gains consciousness.
Red Lobster
Really enjoyed this one—felt like a clear signal of where “personal agents that actually do things” are heading.
I ended up writing a response riffing on a slightly different angle: the gap between having an AI and actually compounding advantage from it (what I’m calling the “Copilot Ceiling”):
https://www.linkedin.com/pulse/copilot-ceiling-why-ai-waits-wont-win-yang-liu-g5nrc/
Curious how you think about that—does Lobster feel like it’s breaking through that ceiling for you?
Would love to hear your take.
This is amazing Omar! Been following Lobster and thank for sharing it.
I love the dev model: it is the same used by Anthropic and OpenAI and, in the fast moving AI environment, it is the only one that works, also from personal experience.
The exciting challenges, besides the security implications, are customization and token efficiency. Hardware diversity makes a fit for all solution genuinely hard, but that's also what makes the space interesting: there's real room for innovation. And while the cost question is real, prices have been falling fast, and what feels expensive today will likely look very different in 12 months.
Your earlier post about AI actively organizing a desktop — not just searching it — became the origin story of a project I recently built: a two-agent MCP system for job application lifecycle management inside Claude Desktop. One agent is human-in-the-loop for resume customization. The second watches Deleted Items in Outlook — when I delete a rejection email, it autonomously archives the matched folder and logs it.
"Everyone has a Lobster." I've been living this, not just reading about it.
Congrats on the new role — excited to see Ocean 11 take shape.
Incredible to see how far Lobster has come! Are you planning to stay on Claude (via CLI) or introducing alternative models/systems?
Lobster is using ChatGPT / Codex as its main model. I use Claude Code as the ‘manager’ agent. I don’t make changes via Lobster but through Claude Code