Agentics Case Study: How Corvus uses AI to remove three hidden taxes on development
Corvus AI is a lean team working in chemical manufacturing. They are about as "AI enabled" as you could be.
Note: this is our first Agentics newsletter case study! We want to use these as opportunities to talk with teams that are on their AI journey, to get a sense of what works for them in practice. If you or your team has an interesting AI case study, reach out to amol@noriagentic.com.
In a world where code is cheap, true differentiation comes from a simple heuristic: does your product actually serve your customers? Anyone can ship an AI slop app. You’ll only earn usage and growth if your code solves someone’s problems.
The team at Corvus knows this better than most. Corvus is bringing AI tools to chemical manufacturing, an old-school industry where relationships and reputation mean everything, and where the real conversations happen in person, on factory floors in places you’ve never heard of.
From Nikhil, CEO:
Most of the people we work with have been in the industry for 20-30+ years. For us to come in and say ‘we can build something better for you,’ that requires a lot of trust. We are constantly on the road, because that is the only way someone tells you what actually works and what doesn’t.
For a lean team, product velocity is the most important thing. Anything that slows velocity is deadly. But being on the road, in person, is a massive blocker for product. How can you keep your product responsive when you need to be out in the field?
That is why the Corvus team decided to roll out Nori Sessions as a background agent service. Background agents let you run coding agents in the cloud, and interact with them from anywhere. Background agents solve a lot of problems that are built into local coding agents like Claude Code. They work where you work — in Slack, Linear, GitHub. They have a single environment that can be shared across a team, so no more having your sales person fiddling with dependencies. And they are always on, so they can solve problems even when you’re busy.
Here are 3 key bottlenecks that the Corvus team has been solving with Nori.
Tax #1: The tracking tax
The default workflow is: hear a feature request, write a ticket, triage, prioritize, eventually build, ship, close the loop. When most conversations happen in person, getting feature requests and feedback can take days. Worse, the cognitive overhead of tracking and remembering compounds with every customer visit.
With background agents, the loop collapses. Hear it, spin up an agent from your phone, get a PR ready by the time you’re back at the hotel. Again from Nikhil:
We were on a customer onsite, and on Day 1 we got some feedback about our UX and some bugs. Instead of throwing those on the Linear backlog where it would sit, we just spun up 3 different agents while we were on the floor, and had the fixes ready to show on Day 2. Huge trust building moment for our users.
Tax #2: The context tax
Slack is where Corvus lives. Customer feedback gets pasted in, bug reports drop into #alerts, decisions get made in threads, services all pipe signals in. It’s the source of truth for what’s happening right now, before any of it gets formalized into Linear or GitHub or Drive.
Local coding agents can integrate with plenty of tools, but the workflow is backwards. They have to take the context that already exists in Slack (the system of record) and then re-stage it somewhere else that the agent can see. Or you have to have a human in the loop, mechanically pasting in context that the model should just have. The result is that every task starts with re-explaining what the Corvus team already knows.
The background agent lives in Slack directly. The agent reads the thread it’s tagged in, picks up the alert that just fired, sees the customer report someone pasted in an hour ago. No restaging, no rehashing. From Alex, CTO:
We discuss everything in Slack, and it’s very convenient to tag Nori directly there so it has all of our decision making and thinking. We’ve added Skills so it effectively takes this context, along with our broader docs (codebase, Notion, etc.), to build the feature and create a PR.
Tax #3: The toil tax
Every hour the CTO spends chasing a Grafana alert is an hour not spent on the hard product rocks. Context-switching is the real killer. This is why teams have massive backlogs that they never get to. Nori absorbs the investigations, the small bug squashing, the “look into this error” work, so humans stay on the complex features. Big features ship faster because no one is getting yanked off them five times a day.
From Alex:
I was at a client dinner Friday night, saw a bug in prod, had Nori look into it, deploy a fix, and re-run customer accounts all while still at the table. We surface all major bugs and errors from Grafana, Trigger, and Render into Slack, then tag an agent to dig in. The turnaround on small stuff is way faster, and it requires way less mind-space.
The compound effect
Remove all three taxes at once and a four-person team starts to feel like twelve. Corvus averages 20 Nori sessions per day, over 100 per week. The downstream effect is more time on the road, with customers, where the real product insight lives.
From Nikhil:
We get to spend so much more time with these manufacturers. It is so powerful to hear live feedback or requests, while onsite, like a dashboard, and then show them the feature a few hours later.
Alex:
This is just the cutting edge of how AI work is meant to be done. It’s made us leaner and more effective, and our customers have really benefited from the flexibility.
Learn more about Corvus (they’re hiring!) at corvusapp.com.
Learn more about Nori Sessions at norisessions.com.
You can find our previous Agentics post on background agents here, and Davy from Windborne gave a talk about background agents at our first meetup here.
Agentics is the study of how to use and reason about agents. If you are an expert in coding agents, or interested in learning more about agents, join our community slack. More articles here. Learn more about Nori at noriagentic.com.
"Hear it, spin up an agent from your phone, get a PR ready by the time you’re back at the hotel."
I'm curious: are y'all using or seeing any kind of tool to automatically take meeting recordings/transcripts (either from in-person meetings or on calls), extract action items from them, and propose turning them into tickets? Seems like that would remove some friction here, I'm curious if this is already built-in with stuff like Granola, Fireflies, etc.