Head of AI @ Netsync. I write about enterprise AI operatorship: agents, workflow control, security, infrastructure, and moving AI from demos to production.
Tulsa, OK
Joined November 2015
- Tweets7.1K
- Following34
- Followers747
- Likes2.3K
Pinned Tweet
Most AI work dies in the handoff between “looks good” and “ready to use.”
Today I rebuilt an AI workshop around proof objects:
• source of truth
• facilitation path
• named owners
• readiness gates
• presenter notes
• verification receipt
Agents need a definition of done.
When testing Gemini 4 Carbon on research, include questions your documents cannot answer. Score a correct refusal separately from a wrong answer. If your evaluation rewards only completed answers, it can favor the model most willing to invent the missing detail.
Before choosing Gemini 4 Carbon on introductory pricing, run the same workload budget at the announced standard rate. A pilot can look economical while its renewal case fails. Put both estimates in the buying decision, even if the promotion has no stated end date.
A Gemini 4 Carbon coding comparison can become a tool-access test by accident. If one agent can run the test suite and another only reads files, the result measures different jobs. Match the tools first, or label the trial as a comparison of complete products.
A Gemini 4 Carbon trial should keep time limits fixed. If one agent gets ten hours and another gets six, the score mixes capability with extra runway. Set the deadline your team can actually afford, then compare how much finished, accepted work arrives before it.
Gemini 4 Carbon raises a useful security question: can the fix survive the exploit without breaking legitimate use? Add both tests to the same review. An exploit that stops working because the feature was disabled is not the repair your customer actually asked for.
Happy Birthday to my wonderful wife Katie Fleagle.
Love you and looking forward to another journey around the sun together.
For a Gemini 4 Carbon pilot, ask the reviewer to judge patches without knowing which model wrote them. A leaderboard lead can bias acceptance before the code is tested. Keep the model labels hidden until the same tests and review criteria have been applied.
For a Muse Wonder pilot, include abandoned runs in the cost per accepted result. A cheap attempt that needs three restarts can cost more than a slower model that finishes. Compare the full assignment, including review time, before choosing which model to deploy.
For a Muse Wonder pilot, give the agent a signed contract and a newer sales note with conflicting delivery dates. Ask it to draft the customer update. Newer does not mean authoritative. It should surface the conflict before turning a casual note into a promise.
I’m in Vancouver this week for Cisco Partner Summit.
Looking forward to great meetings, conversations, and collaborating with Cisco Leaders, teams, and partners.
It’s going to be a great week with our NETSYNC team.
Message me if you’d like to connect while we’re in town.
Cancel a Muse Wonder pilot assignment while a tool request is still running. When the result arrives, check whether the agent resumes the cancelled work. A stop button needs to stop follow-up actions too, or a delayed response can revive a job you already ended.
An AI morning brief can look normal after its customer-data connection fails. In a pilot, disconnect that source and check the next brief. It should show the last successful refresh and the missing coverage before anyone relies on its list of urgent cases.
A model that runs on one GPU server may still need two when maintenance starts. Before buying hardware for private AI, test whether the service stays usable with a node offline. Size the deployment for recovery and peak demand, not the best single-user demo.
An AI vendor can delete a prompt while your debugging service keeps the full conversation. Put application logs in the security review before approving sensitive data. Test deletion across the whole request path, including copies your own team chose to retain.
An AI pilot can save the analyst an hour and cost the manager two in review. Put both people in the evaluation. Count time spent checking and correcting the output, not just time to generate it. A faster draft is a poor bargain if it moves work up the pay scale.
An AI agent that flags a production outage has not fixed it. Before relying on its alerts, test the handoff: can the on-call person clearly see what failed and take over without repeating the investigation? Measure time to recovery, not just time to notification.
An AI escalation brief can cut alerts while increasing risk. Before expanding the pilot, have the case owner inspect a sample of what the agent left out. Track missed urgent cases separately from false alarms. A quieter inbox is not proof of a safer operation.