We built the benchmarks that catch agents failing.
Now we deploy the ones that pass.
Silverstream maps how work actually runs inside your company, tests agents on a clone of your stack, and puts them into production. First agent live in about ten days. Humans stay in the loop. You pay when the job is done.
BACKED BY
Gradient Ventures, founded as Google's AI fund
AUTHORS OF
CUBE, the open agent-benchmark standard
FEATURED AT
NVIDIA GTC 2026, in the ServiceNow Research keynote
IN PRODUCTION
First agent live in about ten days
WHAT WE MEASURED
Most “AI in the enterprise” is a copilot that's idle 90% of the time.
We co-wrote the benchmarks the field uses to test web and computer-use agents. A copilot makes a person about 20% faster, and it's active for roughly a tenth of the working day. An agent that completes the task end to end gives you 100x throughput. The gap isn't model capability. It's the last 20% of a task, where an agent hits a changed layout, an exception, a login, or a rule nobody wrote down.
That last 20% is where automation projects die. It's also exactly what we've spent years cataloguing.
of the working day a copilot is actually active
100x
throughput from an agent that finishes the task itself
61
distinct ways we've catalogued agents failing, across 7 domains
HOW IT WORKS
Map. Simulate. Automate. First agent in production in about ten days.
The same loop we use to evaluate frontier models, pointed at your workflows. Nothing goes live until it has passed on a copy of your systems.
01
72 HOURS
Map
Our agents connect read-only to the tools your teams already use and build a map of how work actually flows. You get a baseline Conformance report: which processes are ready for an agent, and which aren't yet. Your data does not leave your perimeter.
02
DAY 5
Simulate
We stand up a digital clone of your stack and run agents against it. You see error rate, cost and throughput for each candidate workflow before anything touches production, plus the top five automations ready for review.
03
DAY 10
Automate
The first agent goes live on the workflow that passed. Agents discover new data sources, the map updates itself, and we add two more automations in the first month and scale from there. Every exception is routed to a human and fed back into the next run.
Most vendors quote three to six months and a six-figure integration. Our first Voxy deployment took two weeks.
WHY NOT THE ALTERNATIVES
RPA breaks when the screen changes. DIY means building your own test lab.
RPA
Build it yourself
Silverstream
When the UI changes
RPA
Scripts break. Someone re-records them.
BUILD IT YOURSELF
Your team patches the agent.
SILVERSTREAM
Agents read layout and meaning, not coordinates. They adapt.
Testing before go-live
RPA
Manual QA on the happy path.
BUILD IT YOURSELF
You build the evaluation infrastructure first.
SILVERSTREAM
Tested on a clone of your stack, against the benchmarks we co-wrote.
Exceptions
RPA
Stop the run. Ticket to IT.
BUILD IT YOURSELF
Whatever you wired up.
SILVERSTREAM
Routed to a human, logged, fed back into the next run.
Time to first result
RPA
Weeks per process, then maintenance forever.
BUILD IT YOURSELF
Months of engineering before anything runs.
SILVERSTREAM
Map in 72 hours. First agent in production in about ten days.
What you pay for
RPA
Licences and the people who maintain the scripts.
BUILD IT YOURSELF
Engineers, tokens and attempts.
SILVERSTREAM
Completed jobs. Nothing else.
CASE STUDIES
Agents in production, with the numbers.
FINTECH UNICORN
Officer support, logistics and account sign-ups run by one outsourced team. We deployed just-in-time agents that spin up on demand, running in encrypted enclaves.
−80%
turnaround time
Variable
compute, instead of fixed headcount
VOXY · EDTECH
Outbound sales needed researched, qualified leads across Mexico, Brazil and Italy. Our agents research each company and the reason to reach out; Voxy's reps work the leads.
~5x
reply rate in targeted vertical outreach
7.5%
and 5.9%, against a 1.2% baseline
FRONTIER AI LAB
Pre-release models evaluated by internal teams; regressions found late. We run automated regression detection and periodic evaluation reports.
−80%
researcher time to identify issues
Higher
test coverage
ARTICLES
From the blog.
WHO'S BUILDING THIS
Computer-use researchers and repeat founders.
Manuel Del Verme
CEO
PhD in computer-use agents; co-author of WorkArena and BrowserGym. Founded an RL trading company, acquired by Equilibre.
Simone Baratta
CTO
Distributed-systems PhD. Cloud Security Lead at Google; ran security and updates for the EU Google fleet.
Matteo Avalle
Solutions Architect
Distributed-systems PhD. CTO at Fluentify (acquired); led digital transformation in PE-backed turnarounds.
Omar Younis
Researcher, Reinforcement Learning
Former researcher at Mila. Co-author of Gymnasium, the standard interface for reinforcement-learning environments (NeurIPS 2025).
Pre-seed led by Gradient Ventures, founded as Google's AI fund.
REQUEST A DEMO
Pick one workflow. We'll show you an agent running it in two weeks.
A two-hour discovery call to choose the workflow and the success metric. A tailored demo on your documents and edge cases. A pilot with a simulation report and an ROI estimate before you commit.
We reply within one working day. No newsletter, no sequence.
