The shadow org chart
Within sight, agents within the capability ballpark of junior remote workers will proliferate in companies. A human will be able to ask an agent(s) to e.g. "draft a memo assessing how the HamburgerGen launch will comply with our internal privacy guidelines. Consult Lauren, John, and Craig to get their input." The agent will be able to navigate the internal filebase, identify relevant files, draft, and ping the colleagues for comments, then address them. At first they will likely be clumsy (as are junior remote workers), and later versions will be uncannily good.
It's fun to think through the adaptations that companies will have to make, and the otherworldly new organizational etiquettes that will arise.
Access controls. Often in qualitative, corporate work, finding the right resource and getting access to it is an early obstacle. Agents will be quite good at the finding part, but companies will need to develop the right access tiers or systems. These include categorizations of confidentiality that only certain credentialed agents (perhaps apportioned to more senior employees) can access. Alternatively, agents may need to request access to files as humans currently do.
Communication etiquette. We ask things of colleagues that they can't, don't want to, or don't have the time to give us. If I send an agent knocking on the door of Lauren, who is more senior than me, and she tells them no, she'll then be irked if they pester her again later, and even more irked if two more of my agents clumsily go knocking on her door later. Agents will need to have sensitivity to this, and know the right way to report back to me, or even persist with the ask at times. To an extent, I'll be responsible for the faux pas of my agents.
Stonewalling. In some cases, colleagues may simply not want to interact with pesky little software agents and so may not respond. Likely there will be a social cost of doing so. HR may come knocking on my door if I stonewall my VPs agents, or am rude to them. Is this codified in the employee handbook?
Earnestness. What if exasperated John misdirects an agent to Craig, who then redirects the agent to Carla, who gives the agent faulty information? How are agents configured to be skeptical or truth-seeking through these behaviors, or should they simply report back to their human for direction? Will our agents all be docile, earnest goody two-shoes, or tacticians who help us, within bounds, carry out our goals?
Agent-to-agent conflict. Ok, if I'm sending agents around maybe they are simply interacting by proxy with my colleagues through their agents. Sometimes this may work wonderfully, but other times goals may be in conflict if I instruct my agent to get information from Lauren, and she instructs hers to always stonewall Dan. Does a more senior-aligned agent win in this case, or do they simply escalate back to their humans.
Metawork. A new category of work will emerge that's solely intended to be processed at the agent level. Agents may all convene on a contentious budgeting decision to sketch out recommendations that align with their managers/orgs, and discuss. They send a summary of their discussion to humans who will make a call accordingly.
Continual learning. How do we download enough of our tacit knowledge so our agents can do a decent job on their task? I think it's plausible we just dictate a bunch of additional context to them (political dynamics, etc), have some sort of virtual call with them where they can ask follow up questions. Or, the agent pings you "Lauren is stonewalling me. Should I persist?" in a similar way to override questions in current coding CLIs. I'd think all this could be done somewhat clumsily with near-current capabilities, or much better in a world where we've made some progress on continual learning. Of course, the information individuals give these agents may be wrong, or deceptive, triggering downstream issues.
Agent apportionment. At least in the near term, agents will be finite. More senior, trusted employees would plausibly have access to more agents or more capable agents (e.g. with better access credentials, higher inference budget), as they will have the discretion to deploy them. Even if agent apportionment is not constrained by resources, it may need to practically be constrained for other reasons, like maintaining order.
Thinking about agents as "reporting" to individuals may be wrong. Perhaps they are shared by a team, or org, which comes with complications that rhyme with the above - org-agents stonewalling each other, escalating issues back to leaders.
In any of these cases, it strikes me that these are all bizarro transmutations of issues we already face in the workplace, especially within the context of managers and their reports. Making all of this work may be pretty normal in some ways: a combination of IT workers tweaking permissions and drafting agent documentation/protocols, leaders making resourcing decisions, and HR people establishing new codified rules and arbitrating breaches of unwritten etiquette.
A strange new world in some ways, but I think much of this will quickly feel quotidian!