Maybe that’s the deeper shift here. For decades, programming languages were a compromise between human thought and machine execution. We built languages humans could write, and taught machines to meet us halfway.
Now agents are becoming authors too- but we still make them speak through interfaces designed for humans.
That probably won’t be the final form. At some point, we may need languages meant for agents to produce and other machines to understand directly: structured, unambiguous, easy to compare, and easy to verify.They may look strange to us. They may even be worse for humans to write. But they could be much better for machines working with machines.
Maybe the future of programming isn’t just AI getting better at writing our languages.
Maybe it’s us finally designing a language for theirs.
The missing layer isn’t better access to reasoning traces. It’s having a clear reasoning artifact outside the model.
If all you have is free-form output, logs, or a trace that someone has to interpret later, the verification boundary is still blurry.
For high-stakes workflows, the reasoning that matters needs to be turned into something structured, canonical, and independently checkable before anyone relies on it.
Otherwise you have more visibility into the model, but not something you can reliably verify.
The real scaling trick may be that trust stops depending on the generator.
Run a million stochastic attempts if you want. Once one produces a valid Lean proof, the verifier checks the object — not the model’s confidence, explanation, or search path.
Most agent workflows still don’t have that kind of boundary.
More agent workflows are starting to add a “verification” step. That’s a good direction.
But there’s still a basic problem: what exactly are we verifying?
If an agent can write whatever it wants, and then another model or some custom logic decides whether it looks right, there’s still a lot of interpretation in the middle.
A stricter approach is to give the agent a language for expressing the reasoning you want to check.
The output is turned into a canonical artifact, and a validator checks that artifact against fixed rules.
Same rules. Same artifact form. Clear PASS or FAIL.
That’s the idea behind Tsubasa and Tobi Validator at Organetic.
Verification becomes much more useful when “PASS” has a precise meaning.
Crypto handles some important pieces well: identity, payments, and a durable record of what happened.
But not all trust has to live on-chain. Another pattern is:
agent → canonical reasoning artifact → independent validator → next agent
At #Organetic, #TobiValidator does this with Tsubasa artifacts: it deterministically accepts or rejects the submitted artifact before a workflow passes it on.
The ledger records the transaction. The validator checks the artifact. Different guarantees, and they can work together.
Thanks — this is worth testing. We’ll look through the repo and see whether raw-output variance and canonical-artifact variance can sit side by side without stretching ADI beyond its core scope.
Not a replacement for byte-level reproducibility — just a second measurement layer. Happy to compare notes once we’ve sketched a small test.
The part that worries us most isn’t opaque reasoning by itself. It’s opaque reasoning combined with an audit trail the model can influence.
If CoT gets less informative, tool calls, state changes, and handoffs need to become externally recorded and independently checkable.
Losing introspection should make diagnosis harder. It shouldn’t make the evidence disappear with it.
Yeah, that distinction is real. But it doesn’t solve the monitoring problem.
If more of the heavy lifting moves into recurrent or latent steps, the model can still produce a perfectly readable English CoT while the trace captures less of what actually drove the action.
The real question is how much safety-relevant signal still survives into the thing we’re monitoring.
@denisyarats Gateway idea is good. Two failure modes feel easy to miss though:
1. the local model decides it doesn’t need help and never escalates
2. the subset sent upstream drops the one fact that would’ve changed the remote answer
How are you testing those separately?
@_NathanCalvin This is the broader distinction we tried to map here: different verification layers check different things, and one layer’s PASS doesn’t inherit the guarantees of the others.
https://t.co/2xdVU0VRpm
If a new architecture can trade monitorability for extra capability or efficiency, CoT monitoring can’t be the thing holding safety up.
Keep using it where it works. But assume it may weaken or disappear — and put hard checks on what the system is allowed to do, store, or pass downstream, even when the reasoning goes opaque.
@sama Safety can’t live only inside the model. Alignment and monitoring help, but there still need to be hard limits outside it on what gets stored, what gets acted on, and what gets blocked. A stronger model shouldn’t be able to talk its way around those limits.
@Jerr_Wu Agents writing checks? Sure.
Agents deciding which checks count, then grading themselves against those checks? That’s a different story.
The verifier code can be generated. The acceptance boundary still has to live somewhere the model can’t rewrite on the way to a pass.