Upwork Just Released the Only Team Shipping Agents Inside ChatGPT
Upwork cut 145 staff on May 7, 2026, including engineers who built Uma AI and the ChatGPT app. Here's how to source them before they re-sign.
On May 7, 2026, Upwork announced Q1 earnings, bragged about Uma AI and a new ChatGPT app integration as the quarter's headline wins, and in the same breath cut roughly 24% of staff. The engineers who shipped its most agent-native product surfaces are now on the open market. If you are building anything agentic, this is one of the cleanest, most time-boxed sourcing windows of the year.
The setup, in numbers
Upwork's restructuring plan reduces total headcount by approximately 24%, around 145 jobs out of a 600-person company, disclosed alongside Q1 calendar 2026 earnings. CEO Hayden Brown told employees: "we will be executing a layoff that impacts approximately a quarter of our workforce." Her framing was sharper: "Two pizza teams are dead." Smaller senior pods plus AI tooling, she argued, now beat mid-size product squads.
The market reaction was brutal. Q1 revenue came in at $195.5 million, up just 1.4% year over year. Forward Q2 guidance landed roughly 6.9% below analyst estimates. The stock dropped 19.3% on the day to $8.54.
What the same earnings release celebrated, almost in the next paragraph, is the more interesting part for anyone hiring.
What this team actually shipped
Skip the LinkedIn-headline version of "AI engineer." The Uma cohort is not a pile of LangChain tutorialists.
Upwork built Uma by explicitly rejecting off-the-shelf LLMs. The team custom-trained models using proprietary marketplace data, hired expert freelancers to create training datasets, generated synthetic data anchored in real platform interactions, and fine-tuned open-source LLMs for hiring workflows. They added AI engineers and researchers from hyperscaler labs over a roughly two-year build starting around 2023. The result, per Upwork's own writeup, was models that substantially outperform generic alternatives on domain-specific tasks while cutting inference costs by up to 10x.
Pre-layoff, Uma was already powering a majority of new client job posts and increasing successful matches on high-value projects by 8%. The Summer 2025 release alone bundled 75+ new innovations. This was a team shipping at high cadence right up to the day they were cut.
Then there is the ChatGPT app, launched April 9, 2026, where Upwork was one of the first OpenAI Apps SDK launch partners. Inside ChatGPT, users describe project needs, discover talent across more than 18 million professionals, 130 categories of work, and 10,000 skills, then hand off to Uma to scope projects, generate contracts, and start work. Built with OpenAI's jobs and certification team. That ChatGPT-native cohort is probably a dozen engineers and PMs. Almost every AI-native B2B startup wants distribution inside ChatGPT. Almost none of them have shipped it.
The combo that's actually rare
The hard problem in agentic products is not the model. It is two-sided trust, payments, dispute handling, and ranking under adversarial pressure. Upwork engineers have shipped agentic actions, autonomous interviews, contract generation, scoping, inside a regulated, money-moving marketplace.
That makes them directly relevant to:
- Agent-native marketplaces: Mercor, Braintrust, Worksome, Contra, Catch+Release, A.Team
- B2B agent platforms: Sierra, Decagon, /dev/agents, vertical agent startups
- Money-movement agents: Stripe-agent, Ramp-agent, Pylon, expense and procurement agents
- OpenAI's own apps and jobs team, which already worked with this cohort
If your product is "GPT-4 with a system prompt," this is overkill. If it is "an agent that does work and moves money," there are maybe a few hundred engineers in the world with directly applicable shipped experience. A meaningful chunk of them just got cut.
Read the layoff carefully before you pitch
The lazy framing is "Upwork cut its best AI people." Don't lead with that. It is wrong, and the engineers know it is wrong.
Brown's "two pizza teams are dead" line, read literally, says senior pods plus AI tooling replace mid-size squads. That implies cuts hit ICs on shipping teams and middle managers more than research leads. Some of the most senior researchers, the people most identified with Uma's IP, are likely still inside. The people on the market are the high-output engineers who actually integrated Uma into product surfaces, the ones who shipped the ChatGPT app, the platform engineers who made fine-tuned models reliable in production. Different pitch, different angle.
Note also that the restructuring is "substantially complete in Q4 2026." This is not a single-day cut. Severance and notice periods will stagger exits. Recruiters who treat this as a 72-hour sprint miss the larger second wave in July through September.
The people on the market are not the headline researchers. They are the engineers who actually shipped fine-tuned models into a money-moving marketplace.
The competing demand problem
Ex-Upwork engineers are not the only senior pool hitting market the same week. Cloudflare cut roughly 1,100 (about 20%) on May 7. Coinbase cut 14% on May 5, around 700 people. PayPal disclosed a phased 20% reduction on May 5. Freshworks cut 11%. BILL flagged up to 30%.
That is several thousand senior engineers competing for inbox attention from the same set of AI-native acquirers. Generic "we saw you got laid off, want to chat?" messages will get filtered by everyone.
The Upwork cohort has the strongest narrative differentiation in this batch, because their product is literally agentic hiring. If you are building an agent-native marketplace, your outreach should name Uma, name the ChatGPT integration, and reference a specific surface they shipped. Generic AI-startup pitches will lose to founders who clearly understand what the engineer actually built.
This is exactly the friction we built Refolk for. You describe the person in plain English ("engineers who shipped fine-tuned LLMs at Upwork on the Uma or ChatGPT app team, currently in the Bay Area, open to staff IC roles"), and get a ranked shortlist drawn across GitHub, LinkedIn, and the open web, including the public signals (conference talks, ZenML case studies, OpenAI launch partner pages) that tell you who actually shipped versus who just had the title.
A 6-week sourcing plan, not a 72-hour sprint
Week 1 to 2: Map, don't message
Most recruiters waste the first week blasting templates. Use it to build the map. There are roughly 145 names. The relevant subset for an agent-native role is probably 30 to 60. Find them.
Public signals worth combing:
- Authors on the ZenML LLMOps case study and any conference talks about Uma's training pipeline
- OpenAI Apps SDK launch partner credits and OpenAI DevDay videos
- GitHub commit history on Upwork-adjacent open source, plus personal repos that show fine-tuning, evals, or RAG work on marketplace-shaped data
- Public posts from late 2025 and early 2026 announcing shipped Uma features
Warm referral nodes, not targets: Dave Bottoms (GM/SVP Product, Upwork Marketplace), Anthony Kappus (GM and COO), Peter Sanborn (CBO). They are public-facing leaders associated with Uma and the ChatGPT integration. They are also likely staying. Treat them as map-builders, not pitches.
For sourcing AI agent engineers at this scale, the bottleneck is rarely contact info. It is judgment about who actually built what. Refolk's plain-English search is what we use internally to triage 600 engineer profiles down to the 40 that match a specific shipped artifact.
Week 3 to 4: Ship the differentiated pitch
When you do reach out, lead with the artifact. Not "saw the news, sorry." Lead with "saw your work on the ChatGPT app integration, the handoff from discovery to Uma scoping was the cleanest implementation of the Apps SDK I've seen, here is what we are building and why your specific stack matters."
If you cannot write that sentence honestly, do not send the email. The engineer will know.
Compensation conversations matter here. The stock dropped 19.3% on May 7, which means a meaningful chunk of unvested equity just lost value, and severance equity grants are unlikely to recover near-term. A clear, fair cash-and-equity offer with explicit refresh logic will land harder than a vague "competitive package."
Week 5 to 6: Catch the second wave
Because exits run through Q4, expect a quieter mid-summer and a second surge in late July and August. Engineers who took the first severance package and traveled in May come back to job hunts in July. Recruiters who declared the window closed after week three miss the bigger half.
What to actually look for on a profile
For marketplace engineer recruiting, the resume keywords lie. Almost everyone now claims "LLM," "agent," and "fine-tuning." For ex-Upwork engineers hiring decisions, weight these signals instead:
- Shipped surfaces that touch money: contract generation, escrow logic, payment-aware ranking
- Public artifacts about evals, not just training: anyone who has shipped agent products at scale has scars about offline-to-online eval gaps
- Adversarial robustness work: marketplace abuse, prompt injection on user-supplied content, dispute classification
- Apps SDK or function-calling-in-production experience, with a real customer surface, not a demo
- Cross-functional fingerprints: PRs into marketing-page repos, design system commits, infra work. Two-pizza-team alums tend to leave fingerprints across the codebase.
For all of these, the open web tells you more than the resume. That is the entire reason we point Refolk at GitHub, LinkedIn, and the open web together: a single LinkedIn title is not enough signal to know whether someone built Uma's contract generator or just sat near it.
The bigger pattern
This is Upwork's third major workforce reduction in three years: 15% in 2023 (~137 jobs), 21% in October 2024 (~160 jobs), now 24% in May 2026 (~145 jobs). Each cut released talent that was, broadly, more agent-fluent than the last. The 2024 cohort built early Uma. The 2026 cohort shipped Uma into ChatGPT. If your hiring process cannot move on this in a real way, the lesson is structural, not incidental: the next agentic-talent dump is already on the calendar somewhere.
The companies that will benefit most from this Upwork layoffs May 2026 cohort are not the ones with the biggest budgets. They are the ones whose founders can write a three-sentence email about a specific product surface the engineer shipped, with an offer that respects the engineer's actual scarcity.
That is a sourcing problem and a sales problem at the same time. Solve both, and you walk away with two or three of the rarest hires of the year.
FAQ
How many engineers from the Uma AI or ChatGPT app team are likely on the market?
Out of approximately 145 cuts, the agent-native subset is probably 30 to 60 engineers, with maybe a dozen having direct ChatGPT app integration experience. Senior research leads are likely retained; the people on the market skew toward high-output ICs and middle managers who shipped product surfaces. Treat the headline number as your funnel top, not your target list.
Is the hiring window really only a few days?
No. Upwork said the restructuring will be "substantially complete in Q4 2026," which means staggered exits with severance and notice periods. Expect a first wave in May and June, a quieter July, and a second wave in late summer as people return from time off and start serious job hunting. Plan for six to twelve weeks, not a 72-hour sprint.
How is this cohort different from the Cloudflare or Coinbase layoffs the same week?
All three are senior pools, but the Upwork cohort is uniquely positioned for agentic products: shipped fine-tuned LLMs, custom training data pipelines, and one of the first OpenAI Apps SDK production launches inside a regulated, money-moving marketplace. If you are building an agent-native marketplace or a B2B agent platform, this is the most directly relevant pool of the week.
What's the right first message to one of these engineers?
Lead with a specific shipped artifact, not the layoff. Reference Uma's training approach, the ChatGPT handoff to Uma, or a Summer 2025 release feature. Explain in one sentence why your product needs that exact stack. Skip "sorry to hear" and skip generic "we are building the future of AI" copy. Engineers in this cohort can spot a templated pitch in three seconds and will rank you accordingly.