Divya van Mahajan

Why Agents Make Every Job a Startup

ai-daily-brief-podcast

Why Agents Make Every Job a Startup

Overview

This episode of The AI Daily Brief (recorded May 3, 2026) presents a “long read / big think” argument by the show’s host (name not explicitly stated) about a paradox emerging in the agentic AI era: rather than reducing workload, AI agents are dramatically expanding the scope of work people feel compelled to pursue, creating psychological and organisational dynamics that closely resemble the experience of founding a startup. The talk argues that understanding this phenomenon is essential for grasping how AI will practically reshape the economy, as opposed to the theoretical narratives that dominate public discourse.

Source video URL not provided.

Prerequisites

  • Familiarity with the distinction between AI assistants (chat/copilot tools) and AI agents (autonomous, multi-step, parallelisable task executors)
  • Basic understanding of knowledge work and how productivity is typically measured in corporate environments
  • Awareness of the current (2025–2026) state of large language models and agentic frameworks
  • General familiarity with startup culture and the psychological experience of founding or joining an early-stage company
  • Optional: acquaintance with the economic concept of the lump of labour fallacy

Main Points

1. The Productivity Paradox: Working More, Not Less

  • The dominant early narrative around generative AI was time-saving: people would finish work faster and reclaim hours.
  • In practice, people with access to agents are working longer, not shorter hours — logging 6 a.m.–10 p.m. days, skipping sleep, breaking personal wellness rules.
  • Public figures cited include Aaron Levy (Box), Bryan Johnson, and Sam Altman, all of whom describe being unable to stop working because the output per hour is so valuable.
  • The revealed preference of AI builders themselves — switching to polyphasic sleep to maximise tool usage — signals a productivity tool so powerful that time itself becomes the bottleneck, not model quality or access.

2. The Infinite Backlog

  • The lump of labour fallacy holds that there is a fixed amount of work; new technology stealing some of it destroys jobs. This is wrong because of what the host calls the infinite backlog.
  • In any expansionary organisation, there is always more that could be done if time and resources were not constraints. Leaders select a tiny slice of this backlog into a roadmap; the rest remains latent.
  • AI assistants gave workers higher leverage within the same unit of time — compressing time but not breaking it. A Friday afternoon still served as a natural stopping point.
  • Agents break this entirely: they can run 24/7, in parallel, replicating the worker’s intentions across multiple simultaneous workstreams. The infinite backlog transforms from a theoretical future to an immediately addressable present.
  • The psychological result: the infinite backlog shifts from a motivating horizon into a contemporary failure — a never-ending slate of unmet, immediate opportunities.

3. The Startup Analogy

  • The only useful analogy for this experience is entrepreneurship: assembling finite resources to address an infinite backlog that others haven’t tackled yet.
  • Founders operate without blueprints, discover dead ends only in retrospect, and must constantly prioritise from infinite options with finite capacity.
  • Startups produce the highest highs in business (creating something from nothing) but also a Kierkegaardian dizziness of freedom — paralysis or anxiety born from unconstrained optionality.
  • This is why most people do not choose to be founders. Agents are now imposing that same psychological condition on ordinary knowledge workers.

4. The Known vs. Unknown Backlog

  • The first phase of engaging with agents is exciting because it addresses the known backlog: things workers always wanted to do but couldn’t — more content output, better analytics, faster iteration.
  • The challenge emerges when agents make the unknown backlog reachable: work that was never mapped, never prioritised, never even articulated.
  • Once the unknown backlog becomes accessible, workers begin to feel they are always doing the wrong thing, or not enough — the classic founder anxiety.

5. New Constraints Replace Old Ones

  • Agents do not eliminate constraints; they replace them with a different set:
    • Judgment: deciding what is worth working on
    • Planning: sequencing tasks and knowing when to start what
    • Coordination: keeping multiple parallel agent workstreams aligned toward a coherent goal
    • Evaluation: verifying that agent outputs are actually correct and useful
    • Cost: compute and energy supply cannot meet effectively unlimited token demand; this will shape the next 18–24 months more than any other factor
    • Absorption: the intended recipients of all this output (markets, customers) have finite capacity to consume it
  • The new bottleneck is human judgment, not typing or execution. This produces a novel burnout: not physical fatigue from doing, but cognitive exhaustion from deciding, context-switching, and verifying.
  • Practically: workers may get 4–5 extremely intense productive hours per day rather than 8–10 ordinary ones.

6. Required Support Structures

  • To unlock the infinite backlog sustainably, entirely new architectures of support are needed around agent deployments:
    • Technical inputs for agents: model access, token allocations, sandboxes, evaluation tooling, permissioned context
    • Human support: tools and practices for prioritisation, sustainable pacing, and embedded technical support for non-developer agent operators
    • Organisational support: new coordination systems across departments, dynamic management styles responsive to emergent opportunity, and mechanisms to propagate successful unlocks from one team to others

7. Emerging New Roles

  • Rather than speculating abstractly that “new jobs will be created,” the host attempts to sketch specific role archetypes already becoming visible:

    Technical & Infrastructure

    • Agent Ops Engineers — keep agent fleets running reliably
    • Context Librarians — curate and maintain what agents know; manage complex permissioning
    • Eval Engineers — design and operate quality gates at scale

    Coordination & Alignment

    • Coordination Architects — design legibility across parallel workstreams
    • Information Pipeline Owners — route signals to the right agents and humans
    • Orchestration Leads — broker conversations across overlapping agentic work

    Strategic & Managerial

    • Experiment Portfolio Managers — fund, scale, merge, and shut down agentic initiatives
    • Agentic Coaches — support knowledge workers navigating founder-condition stress with judgment, pacing, and wellbeing frameworks

  • Aaron Levy (Box) is cited as already hiring for an Agent Engineering role: a highly technical internal FTE who wires up business systems (Salesforce, Workday, Box, etc.) and codifies workflows for agents. A companion Agent Product Manager role on the business side is anticipated.

8. Practical Conversations Organisations Should Be Having Now

The host closes with four categories of actionable discussion:

  1. The backlog itself — What has the organisation always meant to do but couldn’t? What is now reachable? Who is best placed to pursue which parts?
  2. Supporting the people doing the work — Do workers have model access, budgets, sandboxes, cross-functional context? Is the organisation teaching judgment and prioritisation, not just prompting? Is sustainable pacing being designed, or is overwork being rewarded?
  3. Organisational coherence — Does leadership have ambient awareness of what is being built across teams? Are there mechanisms to spread successful experiments?
  4. Leadership and management itself — How does management change when its job shifts from assigning tasks to harnessing emergent backlog unlock? Which people thrive in the founder condition? What new roles are needed that do not yet exist on the org chart?

Key Concepts

  • Infinite Backlog: The always-present, unbounded set of things an organisation or individual could do if time and resources were unconstrained; normally latent, agents make it immediately addressable.
  • Lump of Labour Fallacy: The erroneous belief that there is a fixed total amount of work in an economy, such that productivity gains destroy jobs by consuming that fixed supply.
  • Agentic Era: The current phase of AI development characterised by autonomous agents that can plan, execute multi-step tasks, and run in parallel without continuous human input.
  • Dizziness of Freedom: (After Kierkegaard) The anxiety and paralysis that arises from having unconstrained optionality — knowing you can do anything but not knowing what you should do.
  • Founder Condition: The psychological state of operating without established blueprints, with infinite options and finite resources, producing both high exhilaration and high stress.
  • Agent Ops Engineer: An emerging role responsible for building, maintaining, and governing fleets of AI agents integrated into internal business systems.
  • Context Librarian: A proposed role focused on curating, maintaining, and permissioning the information and memory that agents draw upon.
  • Eval Engineer: A proposed role dedicated to designing quality-assurance gates and evaluation frameworks for agent outputs at scale.
  • Polyphasic Sleep: A sleep schedule using multiple short sleep periods across 24 hours rather than one long period; referenced as a revealed preference of people maximising agent-assisted work time.
  • Token Demand / Compute Constraint: The emerging bottleneck where demand for AI inference (measured in tokens) exceeds the compute and energy infrastructure available to supply it.

Summary

The central argument is that AI agents do not reduce work — they expose and make immediately actionable the infinite backlog of everything an organisation or individual has always wanted to do but lacked the time and resources to pursue. Where AI assistants compressed time, agents effectively replicate the worker in parallel, dissolving the natural stopping points that previously made the workday feel finite. The psychological result is that every knowledge worker begins to experience something structurally identical to being a startup founder: exhilarating capability combined with an anxiety-inducing dizziness of freedom, a sense of always doing the wrong thing or not enough, and a new form of burnout driven by judgment and decision-making rather than physical execution. The old constraints of time and execution capacity are replaced by new constraints — judgment, coordination, evaluation, cost, and market absorption — none of which agents can fully resolve. The appropriate response is not simply to deploy more agents, but to deliberately design new support architectures: technical infrastructure for agents, human support systems for sustainable pacing and prioritisation, and new organisational coordination mechanisms. New roles will emerge to address these needs, and organisations that begin designing for them now will be better positioned than those that wait. The thesis is simple: agents make every job a startup, with all the opportunity and all the stress that entails.