Yuba WangAI Operations

I run a team of AI agents and keep them honest.

AI operations practitioner. I run a multi-agent assistant every day, with audits, memory, security checks and a human in charge.

I spent years lighting and compositing for film and games.

Now I work in AI operations. My assistant is a small team of AI agents. One routes the work. Others plan, build and run it. Separate agents check the risky parts.

I care less about what agents can do. I care more about whether I can trust what they say.

How my assistant works: me, an orchestrator, workers, checkers, a shared memory store and an approval gate. requestone replyapprove / declinerisky actionread contextplan (bigger work)briefresultbriefresultresult to check (selective)verdictwhen risk is presentverdictMeApproval gaterestarts, deletions,public or paid actionsOrchestratorroutes, verifies, closesShared memory storePlannerplans and acceptance checksCoderprogram codeExecutorsetup, files, commandsAuditorpass / failSecurity reviewerwhen risk is present How my assistant works: me, an orchestrator, workers, checkers, a shared memory store and an approval gate. approve /declinerequestone replyrisky actionplan (bigger work)briefresultbriefresultresult to check(selective)verdictwhen risk is presentverdictread contextMeApproval gaterestarts,deletions, publicor paid actionsOrchestratorroutes, verifies, closesShared memory storePlannerplans and acceptancechecksCoderprogram codeExecutorsetup, files, commandsAuditorpass / failSecurity reviewerwhen risk is present
How my assistant works. An orchestrator takes each request and routes it. A planner, coder and executor do the work. An auditor and a security reviewer check it when the risk calls for it. A shared memory store keeps context between sessions. Risky actions stop at a human approval gate.

Case studies