Korva is a scroll of the internet’s best ideas.
Devin's first task: set up MongoDB
The error loop
Cognition's founders knew the routine. An error appears, so you search for it, find an answer on Stack Overflow or ask ChatGPT, try the suggestion, and hit a different error, which you paste in to start the loop again. They were stuck in that loop with their database setup.
Handing it over
At some point they stopped feeding it fixes and told Devin to just go fix it, running whatever commands it needed. It worked.
The question that stuck
It was a single run, and a better-than-average one. Still, it raised a question that kept Scott Wu awake that night: why shouldn't all software and all products be built this way, by saying what you want and letting an agent go do it? The demos in the launch announcement were real runs the team had done themselves, each one a result that startled the people who saw it first.
Why do error messages lead to new errors?
A database setup involves several layers, including the installed version, configuration files, network permissions, and the operating system. An error message usually names only the layer that failed first. Fixing it lets the process move forward to the next layer, which may have its own problem. This is why setup guides often read like a chain of small repairs.
How does an agent run commands for itself?
The agent is given a terminal inside a sandboxed machine, much like the one an engineer uses. It types a command, reads the output, and uses what it sees to choose the next command. Because the machine is isolated, mistakes stay contained. Stack Overflow answers and documentation, which a person would read in a browser, become text the agent can pull in as well.
Why does one successful run prove so little?
Language models produce different outputs on different runs, so a task that succeeds once can fail the next time. Researchers handle this by running a task many times and reporting the share of successes. Software benchmarks such as SWE-bench are scored this way, as a percentage of tasks solved.
More from David Senra
