Zencoder goes beyond coding
Zencoder on Thursday launched Zenflow for Work, its first product for non-coders.
The company started out as a code completion service in the IDE, but today, the company describes itself as an “orchestration layer for AI engineering” with a focus on agents for planning, coding, testing and code reviews. Similar to how Anthropic built Cowork on top of its experience with Claude Code to offer non-developers an easier-to-use gateway into using agentic AI, Zencoder now also wants to use its platform to expand to a wider audience.
As Zencoder CEO and co-founder Andrew Filev tells The New Stack, the original idea here was to extend the Zencoder platform with tools that would help developers get the same kind of speedups they’ve seen from using coding tools to handle some of their more routine day-to-day tasks like writing standup briefs, release notes, and preparing for meetings.
Those are multi-step workflows that bring in data from services like Jira, Linear, Notion, Gmail, Google Docs and others, and that often run on a schedule.
Thanks to the work the team did on its coding agents, the platform for this was already in place, Filev noted. “We took the platform that we built for coders that already works — it’s performant, it’s well scoped — and we extended it to more proactive work. So doing things like working on schedules. For example, instead of you having to prompt it, it can run something daily, like cleaning your backlog, or it can run every week to prepare reports for stakeholders,” he says.
Filev also stressed that the team implemented the ability for the agent to chase long-running goals, like sheparding a pull request through the review processes (and maybe addressing some comments), all the way to shipping it into production.
Since the team was already building into support for more integration, including for non-coding tools, opening up the platform to non-developers was a pretty straightforward move. “This is where we feel it’s a great opportunity to take it outside of the engineering department and apply across different functions.”
It also parallels what Filev did with his last startup, the work management platform Wrike, which was sold to Vista Equity Partners in 2018 for around $800 million and later to Citrix for $2.25 billion in 2021.
“This reminds me a lot of my Wrike journey. First there were bug trackers for engineers and tools like Jira, which were awesome. Engineers are typically very smart, and they build advanced workflows, and customizations, and it’s wonderful,” Filev says. “And then you’re looking at a marketing department, and you’re like: ‘well, marketing departments should also organize their work.’ And then you try to implement Jira, and they’re like: ‘oh my god, we’re never going to use it.’ So they need a much simpler product with better UX — and this is what kind of gave birth to Wrike.”
Filev argues that the tech industry pioneered a lot of powerful capabilities to help with coding, and now it’s time to package those tools and capabilities into services for non-coders as well.
Thanks to its set of integrations (with support for HubSpot coming soon), the company is targeting sales and marketing teams with this new offering, as well as HR and finance groups.
To bring these tools to business users, Zencoder is running the agents in the cloud. That’s because, as Filev notes, for business users, collaboration is extremely important. “That’s one piece that’s missing from a lot of AI tools right now: the ability to not just collaborate with your agent, but collaborate across the team with the agents,” he says.
Bonus: two models are better than one
Zencoder is also quietly launching another new feature that puts tooling around what a lot of developers are already doing manually today: using a flagship model to write the specs for a new feature and another smaller (and cheaper) model to put those specs into practice. As it turns out, in Zencoder’s testing, this often resulted in better code than using a flagship model.
“The results were that Gemini Flash did a better job at implementation than even Opus, because Opus already kind of put its best foot forward in the plan, so then running Opus for implementation didn’t add much,” Filev says.
So this approach doesn’t just save on token cost — which Filev says can be up to 70 percent — but may even result in better code.
Zencoder is now wrapping its tooling around this practice so developers don’t have to manually run one model to write the spec and then choose another to do the actual coding.
