Passionate about enabling better decisions using #AI & #ML @Alteryx. Interested in #Blockchain #Crypto #PublicPolicy https://xitter.cf/t.co/nXfrcYlyeU
London
Joined April 2009
- Tweets4.8K
- Following4K
- Followers710
- Likes1.9K
A useful harness should make failure cheap for the organisation, not just retries cheap for the model. If an agent stops halfway, who owns the unfinished work? Clear handback and evidence of completion may unlock more adoption than another few points on a benchmark.
The interface that survives may be where authority changes hands. Let the agent do the legwork, then make the commitment visible: what changes, who bears the risk, and how it can be undone. Hiding that boundary makes adoption harder, even when the agent gets the task right.
Prashant Chamarty retweeted
Microsoft CEO Satya Nadella said companies should treat powerful AI models as potential insider threats, assume they could be compromised and create an “emergency brake” system to prevent agentic models from going rogue bloomberg.com/news/articles/…
If finance can build its own tools, implementation stops being the scarce service. The partner opportunity shifts to keeping workflows reliable as they change: exceptions, controls and unit economics. Who owns that after go-live matters more than who built the demo.
A fast decision model makes the gate cheap enough to use everywhere. That also makes a bad threshold scalable. Confidence answers "how likely?" Authority answers "may we?" Keep them separate, or better calibration simply automates the wrong permission.
When everyone can rent the same intelligence, advantage moves to the workflow around it. The test for implementation partners: can the customer run and improve that workflow without you? Transfer capability, not dependence.
Risk belongs to the workflow, not the task label. Drafting is low risk until an agent publishes, prices or pays. Decision rights should change at that boundary, with a named human owner accountable for the outcome.
A judge can agree with a false record as confidently as with a true one. For consequential actions, give it evidence from the system of record, outside the acting agent's control. Independence has to cover the evidence path, not just the choice of model.
The 22-day process often waits on decisions, not work. A background agent can finish its part instantly and still join the same approval queue. Map who can decide, what they can delegate and who owns the exceptions, or faster execution just creates a bigger backlog.
A second check is only independent if the agent cannot alter the evidence. Audit records and the screen reviewers see need separate controls. Otherwise accountability rests on a view the agent can rewrite.
metr.org/blog/2026-10-06-ai-…
Job-replacement forecasts miss who inherits the exceptions. Automate the easy work and the remaining cases get harder, while fewer people learn from routine cases. The operating question is how to keep judgement developing as throughput rises.
If defensive agents inherit the same credentials as the agents they monitor, one compromised identity can disable both. The security budget has to buy separation of authority, not just more detection. Otherwise the watchdog shares the house keys.
Enterprise inverts this. Buyers won't want one shared network holding their identity and context, they will want agents that interoperate. The moat moves to governance: who can prove what an agent was allowed to do. Trust becomes the network effect.
The open question is the business side of the handshake. A protocol tells an agent how to ask, not what the company is willing to authorise. Enterprises will want scoped, revocable, logged permissions per agent before they expose anything real.
Evals matter most at change events. A model upgrade is a procurement and risk event in a regulated firm, so the eval set is what lets a bank say yes to the upgrade in days, not quarters.
Trajectory graders are the underrated one. When a partner delivers the agent and the client runs it, the trajectory is the contract: it shows the process was followed, not just that the answer read well.
The ad point cuts deeper for enterprise. Agents that skip the sponsored page break the business model of every marketplace, so expect gated agent APIs with pricing, not open access.
@ScottTaylor That is the failure to design for. An exception that carries no scope or expiry gets copied by the next case. So every approval should record its boundary, and the check should reject anything outside it.
Output is cheap, so review is the limit. A model's own explanation is tuned to convince, so it cannot be the check. Put a ledger of settled, assumed and open items, and a checker, outside the model. @karpathy prashantchamarty.io/essays/t…
Speeding up one step just moves the queue to the next approval. AI does not remove the bottleneck, it relocates it to wherever decisions wait. The gain goes to whoever redesigns that gate. Everyone else gets faster at waiting. @emollick
