The job
Legacy software is the monster under every big company's bed. Old systems running on old tech, held together by the two engineers who still understand them. Rewriting that stuff takes months, costs a fortune, and nobody volunteers. Odyssey Logistics, a transportation company, needed to rebuild a legacy forms system into a cloud-native platform on a fixed budget and a tight deadline. A normal rewrite wasn't possible. So they tried something that would've sounded crazy two years ago: they put an AI software engineer on the team, gave it real production work, and measured whether it delivered.
The agent
Devin is Cognition's autonomous AI software engineer. Unlike autocomplete tools that suggest the next line, Devin takes a whole task — "rebuild these legacy forms as cloud apps with tests, pipelines, and docs" — and plans, writes, tests, and fixes the code itself. At Odyssey Logistics, a seven-person Cognizant team worked alongside Devin on the rebuild. Every change still went through human review and sign-off by a technical lead before merging, so it wasn't unsupervised. Think of Devin as a tireless junior engineer who never sleeps, paired with a human reviewer who checks everything.
Results
The numbers came from a joint Cognizant-Cognition announcement in September 2026: delivery throughput rose by about a third, and Odyssey recorded a 37% net cost saving versus the conventional approach. The new system is live in production, and the program continues. Odyssey's CIO, Maneet Singh, publicly vouched for the outcome. Devin is also deployed engineering-wide at Hippo Holdings, the insurance company, where it helps build software across fifty states with different regulations (announced April 2026). Both results come from the companies involved, so take them as vendor-and-customer reported — but the named CIO quote is a strong signal this isn't vaporware.
The catch
Here's what the press releases don't headline: every single change got human review and sign-off. Devin doesn't ship to production on its own, and it shouldn't. Coding agents are awesome at well-defined, repetitive engineering work — migrations, test generation, boilerplate. They're still risky on architecture decisions, security-sensitive code, and anything where the requirements are fuzzy. The review bottleneck is the real limit: humans become the QA layer, and if the team stops reviewing carefully, that's when bad code slips through. Also, you need serious process discipline to deploy this. A chaotic team plus an autonomous coder is a recipe for a mess nobody understands.
Who it's for — who should skip
If your engineering team is drowning in well-defined backlog work — legacy migrations, test coverage, integrations — a coding agent like Devin is worth a serious pilot, with humans as reviewers. Skip it if you think it replaces senior engineers. It doesn't. And skip it if your team has weak code review habits, because an AI coder will just generate problems faster than you can catch them. The winning setup is human judgment on what to build, AI muscle on the building.