Many believe that judgment and taste are traits of the most powerful engineers in the age of artificial intelligence.
David Yanacek, senior principal engineer at AWS, believes that the ideal developer has a few additional qualities.
He told Business Insider that today’s best engineers are also “paranoid and lazy.” He says this is a good combination when working with an agency.
“Paranoid means I’m very sensitive to what might happen if I allow an agent to do something without supervision,” said Janáček, who studies agent artificial intelligence and security.
Laziness, he said, meant he didn’t want to perform repetitive tasks manually.
new workflow
While artificial intelligence has taken over most of the coding tasks, Janáček said it’s not the main bottleneck. In addition to writing the actual code, engineers used to spend days waiting for manual testing, code approvals, and design adjustments. These processes require constant waiting for review by teammates.
Now, human intervention is only used for design discussions and complex reviews, while agents handle most other work in a loop.
These cycles involve generating code, running tests, reworking issues, and preparing updates that are then merged and deployed. While a loop can run for several hours, Yanacek can move to other projects or add other features to the product in the meantime.
Set the right guardrails
To make agent workflows work efficiently and save time, agents need to operate autonomously, Yanacek said. So he spent a lot of time deciding which tools agents could use without supervision, which actions required approval, and where those restrictions should be enforced.
To achieve this goal, Janácek ensured that his “automatic trust” files were sound. This is a set of rules that specifies which tools and instructions an agent can use independently and which require human sign-off.
“I always think about automatically trusting what’s in the file because that’s how you increase productivity: giving the agent the ability to do things without having to micromanage every command it runs,” Yanacek said.
His goal is to empower a group of agents to make good decisions without constant review and approval. This requires clear policies on what agents should and should not do.
“My biggest concern is artificial intelligence safety,” Janáček said. “I sleep well because I do this.”
He said engineers need to have a level of paranoia and laziness so that they are alert to problems that can arise when agents act without supervision. However, they also need to build systems that are reliable enough so that they don’t have to oversee every operation.
The ultimate goal, he said, is to do enough work up front so that agents are trustworthy enough to operate without supervision, allowing the tedious parts to be automated.