I’m of the opinion that we need new agent harnesses that have all these modern features while being easy to apply control protocols to and run our evals on, without there being any ways for the model to hide what it’s doing due to some structural issue with the harness. Control is clearly a bit of an afterthought for Claude Code and Codex, even if their built-in control mechanisms are working well right now.
Also, do you have any examples of factored cognition being used by a misbehaving/red team agent successfully? I am looking into studying factored cognition as a control protocol with the expectation that it will make the monitor’s job easier, even when the orchestrating agent and the subagents are untrusted.
(Also, good post, and I broadly agree with everything)
I’m of the opinion that we need new agent harnesses that have all these modern features while being easy to apply control protocols to and run our evals on, without there being any ways for the model to hide what it’s doing due to some structural issue with the harness. Control is clearly a bit of an afterthought for Claude Code and Codex, even if their built-in control mechanisms are working well right now.
Also, do you have any examples of factored cognition being used by a misbehaving/red team agent successfully? I am looking into studying factored cognition as a control protocol with the expectation that it will make the monitor’s job easier, even when the orchestrating agent and the subagents are untrusted.
(Also, good post, and I broadly agree with everything)