DEV Community: Sam Yang

What the Board Asks Me About AI Agents

Sam Yang — Thu, 08 Oct 2026 15:14:19 +0000

I sit in the meetings your architecture decisions eventually reach. Here are the questions that actually get asked about agents at the board level. These are worth knowing, because sooner or later they roll downhill.

"What happens when it's wrong?" Not if. When. The answer they want is a blast-radius answer: what the agent can reach, what it can't, and what record exists. "We review the prompts" is a career limiting answer.

"Why is the AI line item growing faster than usage?" This is the rework question in disguise. If failures re-run completed work, cost grows with instability, not with customers. Boards notice curves that bend the wrong way.

"Are we locked into a single vendor?" They mean the model, the framework, and the cloud. The honest answer is exit cost: what you would have to rewrite to leave, and whether that reliability layer is portable or welded to one stack.

"Who else runs this?" References beat benchmarks at this level. Public case studies are worth more to your internal pitch than any throughput number.

The board will not ask granular questions about the model. They will, however, check to see if you have a thoughtful answer for blast-radius, vendor lock-in and rising costs.

Build strategically and the board meeting goe smoothly.

To better prep for board meetings, review the economics behind running AI at scale: https://www.diagrid.io/blog/economics-of-running-ai-agents-at-scale

The CFO Reads Your Architecture Diagram

Sam Yang — Thu, 01 Oct 2026 07:00:00 +0000

Not literally. But every box you draw becomes a number on a spreadsheet within a quarter, and with AI agents the conversion is harsher than anything I’ve seen in more than two decades of SaaS/Software.

Here’s the translation list:

No durable execution → rework. Crash at step eight of ten, re-run from the top, pay for seven steps twice. Multiply by your deploy frequency. That’s a real line in the token bill: the silent tax.

Side effects with no completion record → reconciliation. If the runtime can't tell whether the refund step finished before the crash, it either skips it or fires it twice. Either way, someone in finance cleans it up by hand, and their time costs more than your tokens.

No execution record → audit overhead. Every compliance question becomes an engineering archaeology project billed by the hour.

Framework lock-in → switching cost. The framework you commit everything to today has a real chance of being the one you migrate off next year.

The metric that fixes the conversation: cost per completed task, not cost per token. Cheap tokens on an unreliable runtime lose to expensive tokens on a durable one, because durable workflows only pay for work once while recovery replays the journal instead of re-running the model.

Quick illustration: a 10-step agent, roughly 1k tokens per step. It crashes at step eight, crashes again on the retry, then finishes on the third attempt. Restarting from the top re-pays steps one through seven each time, so that's 14k tokens of rework on a 10k-token task. Your cost per completed task is 2.4x what the happy path suggests. With durable resume, steps one through seven are paid for once, and the task costs roughly what it should.

Now scale it to a fleet. If 15–30% of workflows fail, as is common in an early deployment, restart-from-zero can add up to that much again to your model bill, and it never shows up as its own line item.

When your architecture makes that metric flat and predictable, budget meetings become short. When it doesn’t, they become frequent.

The full math, worked through with real failure rates: How Durable Execution Cuts Rework and Token Spend and The Economics of Running AI Agents at Scale.

Agent Washing Is the New Cloud Washing

Sam Yang — Wed, 30 Sep 2026 19:07:38 +0000

Have you noticed every vendor is now “agentic”?

We've been through this before. Remember a decade ago when every product was “cloud”? Some were. Most just tweaked their logo to make it look more "cloud-y". It’s happening again but this time the word is “agentic” and the stakes are higher because agents touch everything: customer data, systems, money, and more.

When you meet an "agentic" vendor, here are three tests that can help you cut through the BS and expose who is "agent washing" their product.

The kill test
Ask the vendor to kill their own demo mid-task, live, then restart it. Real ai agent orchestration resumes from the exact step that failed, without re-firing the steps that already happened. Agent washing restarts from zero and calls it “self-healing.”

The evidence test
Ask for the record of what the agent just did – not the chat transcript, the actions. Who did it act as, what did it call, in what order, and can the record be verified after the fact? If the answer involves the word “logs,” keep your checkbook closed.

The economics test
Ask what a failure costs. Durable workflows make the answer specific: nothing already completed is ever paid for twice. If the vendor can’t connect reliability to your token bill, they haven’t run this in production – they’ve demoed it.

One more tell, and it’s the big one: a real platform doesn’t make you abandon your current agent framework. The “rip everything out and use ours” pitch is pure vendor lock-in. Keep your framework. Add the reliability layer underneath. (We wrote up how that works: Keep Your Agent Framework, Add Durable Execution.)

Back in the day, cloud washing stopped working when people learned which questions revealed who was legit and who was not. For agentic vendors, you have 3 questions above.

Fore more detail on Why AI Agents Fail in Production, take a look!

What the Business Actually Needs From Your Agent Stack

Sam Yang — Mon, 28 Sep 2026 04:22:32 +0000

As the leader of an AI infrastructure company, I talk to teams every week who are already getting real work done with agents. The prototypes work. The value is clear. What still stands between them and budget approval is not a better model or a prettier graph; it’s a short list of production gaps finance will actually sign off on.

Predictable cost. Agent rework is the silent tax on your token budget. Every crash that re-runs completed steps multiplies spend. Finance can absorb a big number. It can’t absorb a number that changes for no visible reason. Durable workflows make cost track work done, not failures survived.
Evidence. “The agent decided” does not survive an audit, a customer dispute, or a regulator. What survives is a step-by-step execution record produced by the runtime. If you operate anywhere near the EU AI Act, that record should be signed.
Recovery. Work in flight has to survive deploys. If your answer to “what happens to the 40 tasks running during a rollout?” is “they start over,” the business case has a hole in it.
No lock-in. Frameworks will change — yours probably already did once this year. The reliability layer shouldn’t change with them. Keep the framework swappable. Keep the execution durable.

That’s it. Notice what’s not on the list: which model you picked, which framework you prefer, how elegant the graph looks. Nobody upstairs cares. They care that production agents finish what they start and can prove it.

If your stack delivers these four, the budget conversation gets very easy. The checklist version lives here: what makes an agent production-ready, and the money math here: the economics of running AI agents at scale.

Go build something cool. Go get your budget.