Entries for March 24, 2026
The MCP versus CLI argument should be reframed as Computer vs No-computer argument I personally get the dunk on MCP. It didn't work last year, with earlier models. Then we saw CLIs perform much better with the same models. And giving access to bash was much simpler! Models' training then made them better at calling using a shell. CLIs also have native progressive disclosure, due to the way they work But the most important fact doesn't get pronounced enough IMO A key factor was that giving a CLI to a model also means you are giving it an entire COMPUTER The action space of all commands an agent can run on bash is much, much bigger than a few MCP servers One is a Turing machine, and the other one is basically a REST API. Of course the Turing machine is going to be more powerful, depending on what is at the other end of the API By that logic, giving an agent access to bash over MCP versus direct access to bash should have the same level of effectiveness, with optimized prompt engineering and long term training. Because the interfaces are equivalent So the argument is, should we give our agents access to a computer, or not? It depends on the security requirements and the setup which the agent is supposed to run on. If you are co-hosting the agent on the same machine you are working on, then it is safer to use MCP servers, because it limits the attack surface in case of adversarial attacks But if you are willing to give the agent its own physical computer, willing to be mindful about the lethal trifecta and the principle of the least privilege, giving it shell access is much more useful So MCPs win in restricted/local environments, whereas CLIs/shell access win in unrestricted/remote ones Running an agent locally and safely with shell access requires compartmentalization. This is much heavier compared to installing MCP servers locally, which don't need that. So there is a tendency to use MCP servers locally, e.g. in a work setting Cloud agents on the other hand are more likely to ship with a computer. Because they are already isolated = no risk, and because it makes them much more useful. So cloud agents will be using both CLIs and MCP servers, whichever gets the job done!-
I just registered for an .agent domain and joined the .agent community! @dutifulbob will have bob.agent if it passes :) https://t.co/lhK5MQS1sk @agentcommunity_
-
Sep 2021 @lexfridman podcast with Don Knuth, they also talk about OpenAI Codex (code completion model) around 33 minute mark This aged very well https://t.co/O1eTXlHTNC
-
Damn I’m gonna have to switch to teams if it goes like that
-
Codex's long horizon task and instruction following has been the most life-changing AI feature recently It is unlocking the next level of automation for me. I can convert my own heuristics into prompts and multiply my throughput 100x Currently spending some thought on how to orchestrate all this. Below is a flowchart from a triage workflow I am working onImage hidden
-
-
-
Request for memes A funny and quirky edit of historical timeline of the madness that is openclaw with "Chess type beat" or sth equally jazzy/circusy Preferably including its adventure warelay -> clawdis -> clawdbot -> moltbot -> openclaw Including: - its explosion after @4shadowed's discord integration - naming drama, moltbook and people getting oneshotted about AI takeover - @steipete speedrunning everything - andrew tate calling us gay lol - up to Jensen talking about openclaw on stage for 5 minutes straight and other things I am forgetting maybe overlaid with a lobster just keeping climbing the github star graph and breaking it
-
Native support for Codex on OpenClaw I will be using half my codex channels on acp and other half on codex app server for optimum dogfooding