I’m thrilled to announce that Odessia was acquired by OpenAI, and that I will be joining the team to focus on the future of travel.
I founded Odessia less than a year ago with the goal to ship the best AI travel agent in the market, and I believe we achieved that. Over $60M/year of travel was planned by Odessia over the past summer by tens of thousands of early adopters. Using AI to plan and book extraordinary trips is now a reality.
We’ll have more to share on the shape of the future soon – but for now I want to thank my team, Sequoia and all the phenomenal angels who supported us, and our numerous partners across the travel industry who took an early bet on a startup pre-launch.
How we built Odessia
We got started at a time where agents started really working, and we went all-in.
I previously founded Sonder, taking it from 0 to $600M in revenues which took over a thousand people and a decade. This time, my approach with Odessia was radically different. We burned the old playbooks and radicalized ourselves to truly embrace the new agentic era.
Here’s how we did it.
1. Agents everywhere for everything
With all the context, honed skills, detailed voice prompts and frontier models, agents can multiply productivity and at times one-shot entire projects. And we’re not only talking about software engineering tasks. Strategy, partnerships, design, marketing, customer support, legal, finance — all of it.
My personal usage of agents skyrocketed 25x through 2026. As the founder/CEO, I only had one goal in mind: building a great company fast. As my individual usage surged the ROI from intelligence kept increasing.
2. Context maxxing
Agents need secure access to all the same information an employee has access to. This means MCP or CLI access to every meeting recording, Slack, email, Drive, and all software tools used across the company. This means dozens of plugins/MCPs/CLIs. A single run will retrieve meeting transcripts, review a presentation in drive, look at email exchanges, Slack messages, user feedback, and tap into a few systems of record. The whole is greater than the sum of the parts.
We also created clean shared documentation for every decision for agents to soak up all needed strategy context. Agents maintained this knowledge base and artifact history.
3. Voice prompts & skills
Agents still need a lot of correction to produce excellent work. But writing prompts is too slow. Voice usually goes 3-4x faster. We prefer the methodology where you initially do a brain dump, even expressing your chain of thought to the agent as your thinking evolves. No need to meticulously handcraft prompts.
Before doing pretty much anything manually, we first kicked things off with an agent. As the year progressed, the quality of work kept surprising us. For example, our product launch video was created with an agent. Agents evaluated potential partner APIs by benchmarking their performance and putting together spreadsheets and slide decks with the results.
Repeated corrections were turned into skills that get proactively loaded when relevant.
4. Iterations & frontier models
Tackling a task manually may have taken say 40 hours to do. With agents powered by frontier models even with several rounds of iterations and feedback, you’ve completed the task in 2 hours. 20x productivity gain. The cost of running the agent is a rounding error. The ROI is so good, that the rational thing to do is to go unlimited. We saw productivity accelerate dramatically when we removed usage limits.
The bottleneck ends up being your capacity to consume agent outputs and drive decisions. To process outputs faster, we increasingly asked agents to create one-off artifacts that are more visual. And as agent verification increasingly improved, more work could be done and corrected before human review, e.g. by having another model critique and improve the output in a loop.
Knowledge work is rarely binary which is why frontier models win. You want high performance software, beautiful designs, error free financial models, rock solid analyses, etc. But you also don’t want to have to babysit the model at every step along the way. You want to send high level goals and let the agents rip with as few turns as possible. Your time is the real bottleneck. The best models allow you to reach quality with fewer iterative loops, and thus often more cheaply than models that cost less per token. Cheaper models, however you prompt them, often fail at getting you to clear the quality bar and you end up starting over again manually – a huge waste. Minimizing iterations is far more valuable than trying to save on per-token costs. A contrarian rule of thumb for 2026 has been that the more expensive a frontier model is per token, the higher its ROI.
Note: open source is often pitched as cheaper, but even without considering time/iteration costs and failed outcomes, closed models dominate the entire pareto frontier at any level of intelligence.
5. Lead a team of agents
The workday goes like this: wake up to what the agents did overnight. Voice prompt corrections. Launch a new agent. It takes several minutes to run, but you don’t just stare at it. You launch another one. There’s unlimited work to do to make a business better. This requires multiple areas of ownership for each team member, and the capacity to hold many threads simultaneously. So we had daily standups to divide and conquer and hold each other accountable. Gone are the days of single threads or quarterly roadmaps.
The bottleneck should always be you, and never waiting on agents to do work. Everyone is now a manager of agents, and good managers don’t just look over the shoulders of their team as they do work. They direct, then review completed work, give feedback, and iterate until the bar is met.
It’s also critical to have proactive agents that jump in to do work without being manually launched. Daily analytics reports (not just the same dashboard, also de-novo analyses), investigation of every prod error, review of all customer feedback and traces, accounts payables matching, updating Linear after the daily standup, overnight smoke tests. The sky’s the limit – every single bit of enterprise data can be processed intelligently to improve the product, grow the business faster or improve efficiency.
6. A frenetic pace
You’d think higher productivity enabled by agents means less effort from us humans but the opposite is true. You hear about more and more startups doing 6 or 7 days a week and we were no exception. An hour is so precious, because so much more can get done than before. Can’t let the competition beat us to the punch. Jevons’ paradox applied to our own work schedules.
I’ve always loved the concept of “touch it once” for email, a productivity hack where you force yourself to respond to an email the first time you read it, and then never see or think of it again. I started doing this for any product improvement or idea I wanted to explore. I’d just launch an agent to do it – no task manager, no roadmapping, no prioritization or sequencing – just incessant, instant execution and corrections until the work was high quality. The best process is no process. The ROI on planning and process has cratered.
A new chapter
It’s been an absolute blast to build in this new era of superhuman productivity. I’ve been able to dream of some ideas only to see them in the real world within hours or days, not months or years. In hindsight, it will seem crazy to us that we used to run the economy in such a manual, tedious way.
Anyone that’s owned a roadmap the traditional way knows how painful the tradeoffs are, how little can actually get done, and how messy and complicated it is to deliver on simple ideas.
I can’t wait to see the new capabilities that will be unleashed in the years to come, but I’m optimistic it will have noticeable impacts on products and services across the economy. Expect more customer delight. The bells and whistles that could never make their way onto the roadmap will become a competitive imperative.
And I couldn’t be more excited to help build that future as part of one of the defining companies of our time.



