Relevance AI today announced Invent 2.0, the latest version of our powerful, collaborative agent for designing, building, and managing AI agent workforces. Invent works like a solutions engineer, interviewing you, analyzing your documentation to map out processes and understand how to track value. Invent then creates a new, end-to-end agentic business process built from a coordinated team of specialist agents, featuring the tools, integrations, testing, and governance required for enterprise-class performance and reliability.
The promise of agents is closer than you think
Every senior business leader I’ve spoken to in the last two years signed up hoping to eventually achieve one thing. AI that does the work, does it well, and gets better over time. I’m here to tell you that it’s finally closer than most people realise.
What most teams have so far is AI assistants. A person starts the task, steers it, checks it at the end. This is of course genuinely useful. But if a person is still in the loop on every task, it’s hard to achieve major leaps in productivity.
The hard problem is what happens after agents are built
For a long time at Relevance, we believed that the hard problem was building agents. So we built a platform that made building agents a lot easier, and customers built a lot of them.
Then we watched what happened after the build. Building had stopped being the bottleneck. Keeping agents running reliably became the new one. One customer told us that when his agents errored, “re-processing can take a large part of the day.” Almost nobody had set up testing, so problems only surfaced once something had already gone wrong.
At the same time, the way that people work with AI was changing. You tell an assistant what you want, it does the work, you check the result. Nobody learns the tool anymore. They delegate to it. We were still asking customers to learn our platform, and then to keep everything that they built running on their own.
Solving for a new question - How can agents do the full job of building and managing agents?
Everyone in our space is asking “How do we make building agents easier?” We’d already answered that, and it wasn’t enough. The question we asked instead: what would it take for you to hand over the whole job, the building and the keeping it running, and only get pulled back in when it matters?
That’s Invent. You explain how a piece of work gets done today. Invent asks questions until it understands, builds the agents, tests them, and keeps watching after they go live.
Invent inside Relevance AI. You describe the work; it builds the agents, tools and workforces.
“AI that builds AI” is the kind of line I’d have rolled my eyes at from the other side of the table. So here’s what it looks like with real customers.
What customers are doing with Invent
Fully automating GTM workflows with agents
A RevOps lead at a security software company replaced their deal desk with “one or two prompts into Invent.” $74K a year saved. When he reflected on an earlier outbound workforce project that had taken his team weeks to build, he remarked, “Had we had Invent then, it would have been a day’s work, or maybe an hour’s.”
Troubleshooting agents to find silent failures
A builder at a large SaaS company uses it every day. He pastes a failed run into Invent and asks for a diagnosis before a fix. It quantified a silent failure for him: 13 of 15 LinkedIn searches were returning nothing. He then had it audit 100 runs on his most expensive research agent, drop the dead-weight tools and swap models. “It would have been hard for me to know where to start. And Invent could just do this.”
A failing test, and Invent offering to fix it from the results page.
Finally doing the agent testing that drives success
My favourite is a hospitality group that found Relevance and signed up entirely on their own. They built and published their first multi-agent workforce in a single day, then spent a week having Invent write 150 test cases and 100 checks before going live with their CRM. Over half of everything they approved Invent to do involved testing of their agentic workflows. This was a very welcome surprise. We tried for a long time to get builders to adopt testing on their own. Under 1% did. Invent made it the default.
Quality scored on live runs, an alarm when it drops, and Invent reviewing the dashboard.
Agents have gotten pretty good at executing a task, yet are still challenged at knowing when they’ve got it wrong. So Invent sets up the tests, watches quality on live runs, schedules its own check-ins, and only pings you on Slack or email when something needs a decision. This removes the new “manage and review” bottleneck from humans. With Invent, agents can now spot when they’re going off the rails, without someone sitting there watching them.
Invent schedules its own check-in on an agent and comes back only if something needs you.
Invent has safeguards that minimize the impact of “getting things wrong”
The risk with any agent isn’t that it gets something wrong. It’s that a wrong answer lands somewhere important. So we built Invent around a few critical properties that provide necessary guardrails.
- Drafts. Nothing goes live until a person publishes it.
- Versions. Every publish gets a plain-language description of what changed and why. Rolling back takes about thirty seconds.
- Tests. If you have put tests in place, a failing change can’t ship.
- Permissions. Invent can only do what you’re already allowed to do. Admins turn it on project by project, and set how much autonomy it gets, tool by tool.
One of our largest customers has made Invent the only sanctioned way to change production agents.
Every publish is a named, restorable version. The ones Invent made are marked.
Autonomy is set per action: run automatically, or ask first.
Judge for yourself
Invent won’t turn every process into an autonomous agentic workflow overnight. But it’s the first time I’ve seen what leaders were hoping for - AI that does the work, does it well, and gets better over time - can actually become achievable. And it’s available to Relevance AI customers today.
My question to you: what’s the first piece of work you’d hand over? Book a demo, bring that process, and let’s find out together.
