Pradeep Menon (@rpmchrtvec) header
Pradeep Menon (@rpmchrtvec) profile picture

Founder & CEO, Chartered Vectorial | AI-Native Company Builder | AI Technologist and Futurist | Blogger | Public Speaker | Published Author | @chrtvec

Singapore
Joined May 2008
The biggest change for me as an AI-native founder: I don't manage tasks anymore. I manage decisions. I run a team of 13 people and 15 AI team members. Agents prepare, advise and execute defined work. People decide. My job: make the right calls, and make them clear. #AINative
1
1
75
Who checks the checker? Our content runs on maker-checker: one agent writes, a different agent scores it against a rubric. Only exceptions and spot checks reach me. I check it by marking a random sample of what it passed. Two misses in a week and auto-posting switches off.
2
23
If an agent is on your org chart, it needs a named owner. Not the vendor. Not "the AI team". A person who sets its mandate, reads its evals and answers for it when it gets something wrong. An agent without an owner isn't a team member. It's an unmanaged risk.
10
The model is not the moat. Your definition of good is. Anyone can rent the same model tomorrow. What a competitor can't rent is your rubric: what good looks like in each workflow, written down and tested. When a better model ships, your evals tell you whether to switch.
12
Verification is the new bottleneck. Agents produce faster than a small team can evaluate, and errors pile up silently. Build a control plane: evaluations, observability, audit trails, permissions, rollback, fallbacks, human escalation. Execution gets cheap. Assurance doesn't.
24
Lean is not independent. "Ten people" is a payroll fact, not your production system. Pair the org chart with a dependency map: Critical models and services Data and IP ownership Switching options Key contractors Fallbacks Leverage is rented. Resilience must be owned.
12
Three human gates. Everything between them can increasingly be agentic. Intent gate: should we do this? Risk gate: is the downside acceptable? Release gate: is there enough evidence to ship? Put people at the gates, not on every step.
18
Where should AI autonomy stop? Four tests: Reversibility: can we undo it? Assessability: can we measure it? Consequence: how bad is the downside? Exceptions: how often does it go off script? Reversible, measurable work: delegate. Irreversible calls: a human approves.
1
1
15
AI removes doing faster than it removes choosing. If execution speeds up but every decision comes back to me, I've built a faster queue outside my door. The two-week test: if I vanish for two weeks, can we still ship, support customers, collect cash and handle an incident?
13
Satya's sharpest line isn't about models. It's about authority. You can delegate intelligence to a model. You cannot delegate accountability. When an agent acts, the customer, the regulator and the board still look at the company. The CEO question is not which model to trust. It is where authority sits, and who owns the moment work reaches a customer. That is also a growth strategy. The companies that hand agents the most work will be the ones that drew the boundaries first. Clear authority is what lets you say yes faster. At CV, one agent makes each post, a separate agent checks it, and a person owns anything that reaches a customer. The boundary is why we can let them run.
22
We have evals for code. Why not for the whole organization? Look at how we create content now. One agent makes each post. A separate agent checks it against a rubric and scores it. Anything below the bar goes back for another round. Only exceptions reach me. Maker and checker, never the same agent. Evals and loops are the operating system of an AI-native company. The challenge is to lift them from the codebase to every workflow. Code was just the first workflow we learned to test.
3
124
The autonomy ladder: 1 Assist: AI drafts, human performs 2 Execute: AI acts, human approves each action 3 Operate: AI runs a bounded workflow, human handles exceptions 4 Optimize: AI tunes execution within budgets and policies 5 Decide: low consequence, or behind a human gate
1
19
Climb it one rung at a time, and ask for evidence of success, not the agent's own claim of completion. AI can be given authority. It cannot absorb accountability.
6
Autonomy inside. Accountability at the edge. Inside the company, agents draft, check and prepare work on their own. Where work reaches a customer, a person presses the button. It costs a few minutes. It catches mistakes before a customer sees them.
16
Don't ask which job titles you need. Ask which capabilities you can't lose. Four stay inside: 1. Product and domain judgment 2. Customer relationships 3. Architecture and security ownership 4. Accountability for outcomes You can delegate the work. You still own the promise.
17
Anthropic has made Claude Code moddable. A mod is a small file of code that runs inside the agent. It can watch each tool call, rewrite or block it, add commands and redraw the interface. You can ask Claude to write one for you. Builders are already drawing comparisons to Pi, the open-source coding agent built around extensions. The AI-native point: teams can now shape their own agent. Mods run with your permissions, so someone has to decide who writes them, who reviews them and which ones run. Agents execute. People still decide.
1
1
168
Every company needs five functions: Intent: what are we trying to achieve? Context: what do we need to know? Decisions: what should happen next? Execution: how does the work get done? Verification: how do we know it worked? AI takes more of execution. The rest stays yours.
1
1
45
Every workflow we hand to an agent comes with its own evals. Take hiring. Our workflow combines rubrics and evals. The rubric defines what we look for: how fast someone learns to work with AI, their appetite to try new things, and proof of initiative. The evals check that the agent's work holds up against that rubric before a person makes the call. Evals are the guardrails for agentic execution. We build them into the workflow when we agentize it, not after something breaks. No evals, no autonomy.
2
26
A ten-person unicorn is possible. But not with ten people's worth of capability. The innovation isn't cutting headcount. It's separating judgment from execution. Stop asking "How many people work here?" Ask "How much governed capability can we reliably deploy?"
9
Good framework. We start from the same question, framed as a spectrum. Every workflow sits somewhere between more human and more agentic. Where it sits decides what it needs: evals, risk handling, how much human oversight. Routine work moves agentic, with monitoring. Consequential work stays human-led, with agents preparing and a person deciding. The real design work is placing each workflow on that line, then building guardrails to match.
What tasks should AI be trusted with, and when should humans remain involved? Designing those decision rights is one of the defining challenges of the AI era. Researchers at @MIT_CISR have developed a framework to help. Learn more: bit.ly/4zed20M
1
35
The hardest part of running an AI-native company isn't the technology. It's writing things down. Agents can only work inside what you've made explicit: your standards, your process, your way of working. Vague in, vague out. Clarity is the new leadership skill.
1
64