Divya van Mahajan

Why Agents Still Need Humans

ai-daily-brief-podcast

Why Agents Still Need Humans — Study Document

Overview

This episode of the AI Daily Brief (dated 2026-05-24) examines the evolving patterns of human-agent collaboration in an era where AI agents have become practically mainstream. The central thesis is that rather than eliminating human work, AI agents are paradoxically expanding the volume and complexity of work that requires human judgment, expertise, and oversight. The speaker synthesises his own observations with an essay by Dan Shipper (CEO of Every, an AI-native publication/product/consultancy company) titled “After Automation” to build a coherent argument about why agents still need humans—and how the patterns of that collaboration are maturing.

Source video: No URL provided in the submission.

Prerequisites

  • Basic familiarity with AI language models (LLMs) and how they are prompted
  • Understanding of the distinction between assisted AI (prompt → response → human action) and agentic AI (delegated, autonomous task completion)
  • Awareness of tools mentioned: Claude Code, Codex (OpenAI), OpenClaw (open-source agent harness), Intercom/Fin (customer service platform)
  • Familiarity with concepts such as fine-tuning, benchmarks, and token consumption in AI systems
  • General awareness of the 2025–2026 wave of agent frameworks and harnesses

Main Points

1. The Shift from Assisted AI to Agentic AI (Context Setting)

  • As of early 2026, the dominant paradigm has moved from “prompt and wait” to spinning up and managing agents that produce work autonomously.
  • This shift was driven by model capability improvements at the end of 2025 combined with better harnesses (interfaces for managing agents, e.g., Claude Code, Codex).
  • Business models are under pressure: heavy AI users now consume hundreds of millions to billions of tokens individually per month, creating a token shortage—where demand for AI compute exceeds available supply.

2. The Infinite Backlog — A New Type of Overwhelm

  • The speaker’s earlier concept, the infinite backlog, describes how agents remove the natural ceiling on how much work a single person can attempt.
  • In the assisted AI era, work had a practical end because humans were the bottleneck. Agents don’t tire, don’t stop, and are always ready for the next task.
  • The result is not leisure or job elimination but a new psychological pressure: the feeling that everything not yet delegated to an agent is simply work you haven’t assigned yet.
  • This is a distinct and largely unanticipated consequence of agentic AI—neither the utopian (more free time) nor dystopian (job loss) prediction matched the lived reality of advanced users.

3. Dan Shipper’s “After Automation” — The Paradox of More Human Work

  • Every (an AI-native company of ~30 people) uses agents pervasively: Claude Code, Codex, and similar tools across coding, writing, design, and customer service.
  • Despite maximum automation, human workload has increased, not decreased. No mass layoffs in favour of agents; human writers, editors, engineers, and customer service staff are still employed.
  • The nature of work has transformed: managers commit code, engineers talk directly to customers, AI handles 95% of the CEO’s email—but humans remain essential.
  • Shipper’s core argument: AI commoditises the residue of human expertise (i.e., whatever has been made explicit enough to train on), which collapses the value of default model output and simultaneously creates new demand for differentiated, expert human work.

4. Two Modes of Human-Agent Collaboration

  • Mode 1 — Agents as Employees: Agents are delegated work and go off to produce outputs asynchronously without the human in the loop. Two sub-types:
    • Co-worker agents: Tagged in Slack (e.g., “Andy” at Every, who collects story ideas and produces newsletter digests).
    • Embedded agents: Live inside a specific workflow (e.g., “Fin” handling customer service chat and email).
  • Mode 2 — Human-Agent Collaboration in Work Operating Systems: Tools like Codex and Claude Code function as shared operating environments where humans and multiple agents work simultaneously on complex, original tasks. This mode is described as stranger but more important.
  • Both modes still require a human in the loop for quality, direction, and judgment.

5. The Human Sandwich Model

  • Coined by an Every employee, the human sandwich describes the structure of effective human-agent collaboration:
    1. Human sets the frame: defines the goal, constraints, and success criteria.
    2. AI collapses the task: drafts, searches, codes, summarises, compares.
    3. Human judges and extends: evaluates quality, determines next steps, decides where the work belongs.
  • This structure applies not just to coding but increasingly to writing, email, research, and broader knowledge work.
  • The implication: agents are not autonomous replacements; they are middle-layer executors bookended by human intelligence.

6. Why Abundance Creates Demand for Difference (The Commoditisation Loop)

  • LLMs are trained on the “exhaust” of past human competence—code, prose, support tickets, specs—and make previously rare skills broadly accessible.
  • Cheap competence → rapid adoption → sameness (slop). When everyone uses the same models trained on the same corpus, default output converges to undifferentiated mediocrity.
  • Slop is defined not as any single error but as visible sameness repeated at scale—the predictable output of homogeneous tool use without expert direction.
  • Humans rapidly detect sameness and demand difference. That demand for difference is, structurally, demand for human expert judgment.
  • Key distinction: Current models know about work that has been done. Humans know about what needs to be done right now. Planning, prioritisation, and originality remain human-gated.

7. Team-Based Agents vs. Personal Agents — An Organisational Lesson

  • Early experiments at Every involved each employee having a personal agent (a digital replica of themselves). This failed because:
    • Every time an agent broke, the individual had to fix it themselves.
    • Maintenance burden fell entirely on the individual; not everyone wants to be an agent tinkerer.
    • Personal agents’ value disappears when the employee leaves (continuity problem).
  • The more successful model: shared team agents that serve multiple people whose work overlaps.
    • When capabilities need updating, one person updates the agent and the whole team benefits.
    • Team agents retain company context and act more like a project manager or chief of staff than a private assistant.

8. Evolving Hardware and Workflow Patterns (Harness Maturation)

  • Early agentic setups (e.g., OpenClaw on a Mac Mini with Telegram check-ins and heartbeat loops) prioritised maximum autonomy but produced high token burn and maintenance overhead without proportional output quality.
  • Practitioners like Matt Schumer have moved away from the always-on autonomous OpenClaw model toward more interactive harnesses like Codex.
  • Nick Bauman (OpenAI) describes a multi-device Codex setup (Mac Mini as hub, MacBook and phone as satellite devices) enabling persistent, cross-device agent threads—a shift toward semi-synchronous collaboration.
  • The emerging sweet spot is between:
    • Purely turn-based (prompt → wait → review → repeat), which is too slow.
    • Mega-autonomy (heartbeat-driven, minimal human touch), which produces low-quality or misaligned output.
  • The goal is reduced latency between human guidance and agent execution—working more synchronously without abandoning oversight.

9. Practical Recommendations from the Episode

  • Personal level: Use tools like Codex with voice-based input and steering features to compress the latency between human guidance and agent execution. Reference: the speaker’s “nine tips for getting the most out of Codex” episode.
  • Organisational level: Map overlapping work across team members. Identify tasks that live in the shared space of multiple people’s jobs and build or assign agents to those shared workflows rather than giving each person a separate personal agent.
  • Mental model for AI and employment: Rather than fearing a tipping point where jobs vanish, consider that net agent impact is likely to expand employment in proportion to new categories of work that become possible.

10. Market Signals — Growth Over Efficiency

  • Gartner argues that even with short-term AI-related layoffs, AI will create more jobs than it eliminates beginning around 2028.
  • Case study: Atlassian — announced 10% layoffs in March 2026, stock declined; then reported 29% earnings growth driven by AI-enhanced products, stock rose 29% in a single session.
  • Market analyst Dan Ives argues that companies publicly framing AI as a job-elimination tool are damaging themselves; markets are increasingly rewarding AI-driven growth, not just AI-driven cost cuts.
  • The competitive differentiators in an LLM-commoditised world are: people, engineering talent, and marketing—i.e., human expertise.

Key Concepts

  • Infinite Backlog: The condition created by agents where work is never truly finished—there is always more that could be assigned to an agent, making the total workload feel unbounded.
  • Harness: The interface or framework through which users interact with and manage AI agents (e.g., Claude Code, Codex, OpenClaw).
  • Human Sandwich: A collaboration pattern where humans set the frame and judge/extend the output, with AI collapsing the task in between.
  • Slop: Undifferentiated, mediocre AI output that results from widespread use of the same models without expert human direction; characterised by sameness rather than any single type of error.
  • Token Shortage: A supply-demand imbalance in which total desired AI compute consumption exceeds available compute infrastructure.
  • Agents as Employees (Mode 1): Asynchronous delegation to AI agents that complete defined tasks without continuous human supervision.
  • Work Operating Systems (Mode 2): Collaborative environments (e.g., Codex, Claude Code) where humans and agents work simultaneously on complex tasks in shared workspaces.
  • Human Premium: A framework (developed by the speaker in a separate episode) identifying seven categories of value that do not transfer to AI even when AI can technically perform the task.
  • Heartbeat: A timed self-reminder mechanism used by autonomous agents (especially in OpenClaw) to ensure continuous progress toward goals without human prompting.
  • Semi-synchronous Collaboration: A working mode between fully turn-based (waiting for each AI response) and fully autonomous (minimal human contact), characterised by frequent but not constant human intervention.
  • Shared Team Agent: An agent serving multiple employees whose work overlaps, maintained centrally rather than individually, retaining institutional knowledge independent of any single person.
  • Residue of Human Expertise: The visible, recorded outputs of past human work (code, prose, designs, tickets) on which LLMs are trained—what AI can replicate, and therefore commoditises.

Summary

The central argument of this episode is that the first sustained wave of real-world agentic AI use—stretching across early 2026—has not produced the predicted binary outcomes of either radical leisure or mass job elimination. Instead, practitioners at the frontier of AI-native work (exemplified by Every and its CEO Dan Shipper) report a paradox: the more they automate, the more expert human work there is to do. This paradox is explained by two structural dynamics. First, LLMs commoditise whatever human competence has been recorded and made explicit, which collapses the value of default AI output (“slop”) and simultaneously creates new demand for differentiated, expert human judgment that models—trained on the past—cannot supply for novel, present-moment decisions. Second, agents themselves require ongoing human involvement as bookends: humans must frame work, evaluate outputs, determine next steps, and manage the agents themselves. The practical upshot is that the most effective collaboration pattern is neither purely autonomous agents nor purely turn-based prompting, but a semi-synchronous “human sandwich” model in which humans and agents work closely together in shared environments like Codex and Claude Code. Organisations are also learning that shared team agents outperform personal agents for most use cases, distributing maintenance burden and preserving institutional knowledge. The speaker concludes that companies and individuals who invest in the capability to work with agents—rather than simply deploying agents to replace workers—are the ones most likely to succeed, and that the net employment effect of agentic AI is, on the available evidence, expansionary rather than eliminatory.