Your New Job: Managing a Team of Agents Writing Code, and You Can’t Fire Them

On X, Box CEO Aaron Levie offered a judgment: job descriptions for many roles will change, not just programming. We are moving from “Agents help us do our work a little faster” to “we help Agents finish the work.” The original tweet is here: https://x.com/levie/status/2117134917847093491

The context for this statement is that Agents are AI programs capable of executing multi-step tasks on their own. Levie believes information-intensive work will be affected first: programming, cybersecurity, law, biotech, engineering, and many other roles that process large amounts of information. People in these positions will become managers of a team of Agents rather than individual contributors with a bit of AI added on.

In knowledge work in the past, the main variable you could adjust was hours worked. Assuming talent, collaboration, and strategy remained constant, if you wanted more output, you spent more time. Now that Agents can work in parallel, hours and output are beginning to decouple. In theory, you can spin up enough Agents at the same time to push a goal forward. The sooner individuals and organizations recognize this and reorganize their processes around it, the more likely they are to gain compounding returns.

Uncle Bob Martin relayed a line from the Proton CTO: Your job is not to use Agents to help you write code; your job is to lead Agents and help them write code. Original post: https://x.com/unclebobmartin/status/2116512351491084344

This diagram breaks the process into four stages: doing it yourself; passively letting AI do it; discovering leverage and wanting to use it to do things five times bigger; and finally entering AI plus thinking—planning, prompting, evaluating, iterating, and combining ideas. The truly hard part is the last stage.

In the comments, one user distinguished between “using Agents” and “leading Agents”: using means assigning tasks and accepting the output; leading means guiding, observing, correcting, teaching, and helping them improve. Put plainly, none of the management actions are missing.

Some people also shared results. Pedro Arantes’s line chart shows that in the OneClick Ads project, his weekly merged PR count rose from 30 in the week of September 20, 2026, to 105, and reached 148 from October 18 to 22. The chart marks an inflection point. Original post: https://x.com/arantespp/status/2113337960385126802

His previous approach was to coordinate Agents himself: pick the next task, start a session, explain the context, follow up, review, then switch to the next one. The ceiling was capped by his personal coordination capacity. Only after changing his way of working did the curve start to rise.

But parallelism does not mean it’s easy. One user said you can run ten Agents at once, but you still read their output one at a time. Another user manages six Agents and was still fixing their merge conflicts on Sunday night. Someone else reminded people that many are pushed into the position of “Agent manager” with nothing but a login account and no management training. What’s more awkward is that this manager can’t fire their subordinates.

Management experience does indeed transfer here. If the task description is unclear, Agents will go off track like new employees. One user in the distribution business spends most of their time writing briefs, setting constraints, and reviewing handoffs; the usual reason an Agent fails is that the scope is too vague. Another user said managers still need to understand the work, otherwise they can’t tell a good result from a confident mistake.

A further question also emerged in Levie’s comments: if Agents can manage Agents, will they keep climbing up the org chart and eventually become better Agent managers? There is no answer to this question now, but it is more worth watching than “whether AI can write code.”

In the short term, the scarce skills are very concrete: writing clearly what “done” looks like, setting good boundaries, reviewing handoffs, and being accountable for the result you finally sign off on. Management skills are spreading from manager roles to every knowledge worker. Some people like management, some don’t; the next gap may emerge right here.