First OpenAI Killed Sora for Agents. Now Perplexity Killed Search for Agents. See the Pattern?
Something interesting happened in the first two weeks of April 2026. Two of the most important AI companies on the planet sacrificed their flagship products - and placed the same bet.
OpenAI killed Sora. Not because it failed, but because Sam Altman decided the compute was better spent on what he called "the next generation of automated researchers and companies." Translation: autonomous AI agents.
Then, on April 8, the Financial Times reported that Perplexity's revenue surged 50% in a single month after pivoting away from search - the thing everyone thought would make them the next Google - and going all-in on AI agents. ARR hit $450 million. 100 million monthly active users. And they're just getting started.
Two companies. Two flagship products sacrificed. Same destination.
If you're a founder or CTO building with AI, this isn't a trend to watch. It's a signal to act on. But here's the thing nobody in the hype cycle wants to talk about: the infrastructure to actually evaluate and monitor these agents barely exists.
AI Agent Evaluation Is the Gap Nobody Wants to Admit
When Perplexity was a search engine, measuring success was straightforward. Relevance. Speed. User satisfaction scores. The same playbook Google perfected decades ago.
Agents don't work like that.
An AI agent doesn't return a list of links. It makes decisions. It takes actions. It books things, writes things, moves money, sends messages - on behalf of your users. The failure modes aren't "bad result." They're "wrong action, taken confidently, with no one watching."
And yet, most companies deploying AI agents are evaluating them with the same tools they'd use for a REST API. Uptime. Latency. Error rates. That's like evaluating a new hire by checking if they showed up to the office.
AI agent evaluation needs to answer harder questions. Did the agent accomplish the goal? Did it take the right steps to get there? Did it drift from its intended behavior over time? Did it make a judgment call that a human would disagree with?
Most teams can't answer any of those. That should worry you a lot more than which model you're using.
AI Agent Performance Tracking for the Agent Economy
Here's what Perplexity's numbers actually tell us. The agent economy is real. $450 million in ARR, growing 50% month-over-month - that's not a research project. That's a market.
And every market that grows this fast develops the same problem: quality control at scale.
Perplexity CEO Aravind Srinivas went on the All-In Podcast recently and said "the reality is most people don't enjoy their jobs" - defending the idea that AI agents replacing human work is a net positive. The internet roasted him for it. But buried in the backlash is a real question that founders should be asking themselves.
If AI agents are replacing human workflows, where's the performance review?
Human employees get KPIs, quarterly reviews, feedback loops, and managers watching their output. AI agents get... deployed. Maybe some logs. Maybe a Slack alert if something catches fire. That's not AI agent performance tracking. That's hope.
The companies that win in the agent economy won't be the ones who ship the most agents. They'll be the ones who know - with data, in real time - whether those agents are actually doing what they're supposed to do.
AI Agent Analytics and Monitoring: The Missing Layer
Think about what happened with cloud computing. First came AWS and the raw capability to spin up infrastructure. Then, years later, came the observability stack - Datadog, New Relic, PagerDuty - that made cloud computing actually manageable at scale.
We're at that same inflection point with AI agents. The capability is here. OpenAI is building it. Perplexity is monetizing it. Every YC batch is shipping agents. But the observability layer - the AI agent monitoring and analytics infrastructure - is still in its infancy.
This matters because agents fail differently than traditional software. They don't crash. They drift. They hallucinate. They take confidently wrong actions. They solve the wrong problem elegantly. And without proper monitoring, you won't know until a customer tells you - or worse, leaves without telling you.
The pattern from the cloud era is clear. Capability always comes first. Observability comes second. But the companies that build the observability layer capture an outsized share of the value, because everyone who builds on the capability layer eventually needs it.
That's exactly where we are with AI agents right now.
What This Means If You're Building with Agents
OpenAI and Perplexity aren't making small bets. Altman killed a product burning a million dollars a day. Perplexity walked away from the "Google killer" narrative that raised them billions. These are irreversible, high-conviction moves.
If you're a founder building AI agents - or a CTO deploying them inside your org - the question isn't whether agents are the future. That debate ended this month.
The question is: do you know if your agents are working?
Not "are they running." Not "are they responding." Are they accomplishing their goals? Are they making good decisions? Are they getting better or worse over time?
If you can't answer those questions today, you're flying blind in a market that's accelerating faster than anyone predicted. And the gap between "deployed an agent" and "deployed an agent that provably works" is where the next wave of enterprise value will be created.
Tommbo is building the evaluation and monitoring layer for AI agents - so you can actually prove yours work.
Sources:
- Perplexity's Shift to AI Agents Boosts Revenue 50% - PYMNTS
- Perplexity revenue surges 50% - Financial Times via MarketScreener
- Sam Altman on killing Sora - BusinessToday
- Aravind Srinivas on AI layoffs - Fortune
*The companies that defined the cloud era didn't build the cloud. They built the dashboard. The agent era will be no different.*