Mapping the path to AGI

The agent continuum: mapping the path to AGI

The journey towards artificial general intelligence (AGI) isn't a single leap but a spectrum of increasingly autonomous agents. We're currently seeing coding agents like Cursor and Windsurf hovering around the interactive agent stage, inching towards becoming more ambient with features like background agents. Frameworks like LangGraph and Langchain were early pioneers in this ambient direction, with now OpenAI Codex and Jules beginning to take the coding space. Understanding this progression is key to grasping where we are and where we're headed.

Each step up this ladder of autonomy involves automating tasks traditionally handled by humans. Let's break down these agent categories and how they tackle the fundamental phases of work: explore, plan, execute, and review.

1. Eager agents (human-always-in-the-loop)

These agents are essentially advanced tools requiring constant human guidance and step-by-step validation. They operate in real-time, with the human being an integral part of every micro-decision.

• Explore: The human performs almost all exploration, feeding specific data points or queries to the agent. The agent might assist in fetching or displaying information as directed, but doesn't initiate discovery.
• Plan: Planning is entirely a human endeavor. The human defines the tasks, sub-tasks, and the sequence of operations.
• Execute: The agent executes very small, discrete tasks under direct human command. Think of a simple code completion tool that only suggests the next few tokens based on immediate context.
• Review: Real-time human validation is constant. Every output is scrutinized before proceeding.

This stage is analogous to micromanaging an intern on their very first day.

2. Interactive agents (human-in-the-loop)

This is where current coding agents largely reside. They exhibit some autonomy but require frequent check-ins and operate in a somewhat real-time fashion. The human is still in the loop, but not for every single atomic operation.

• Explore: Humans still lead exploration but can delegate broader information-gathering tasks. For example, asking an agent to "find all instances of this function call in the codebase." The agent explores within defined boundaries.
• Plan: Humans set the high-level goals, and the agent might propose a sequence of steps or a basic plan for a specific, well-defined task (e.g., "refactor this function to improve readability"). The human reviews and approves this plan.
• Execute: The agent can execute more complex sequences of actions, like generating a block of code or attempting to fix a simple bug based on a description. The human intervenes if the agent goes off track or needs clarification.
• Review: Check-ins are frequent. The human reviews chunks of work, provides feedback, and course-corrects. This is like a senior engineer guiding a junior developer, reviewing their pull requests and providing feedback.

3. Ambient agents (human-on-the-loop)

These agents operate semi-autonomously and have lower latency requirements. They can work for extended periods (hours) before needing a human check-in. LangGraph and similar frameworks are pushing into this territory.

• Explore: Agents can conduct more extensive, open-ended exploration based on broader goals. For instance, "research existing solutions for X and summarize their pros and cons." The human sets the direction but isn't involved in the minutiae of the search.
• Plan: Agents can generate more comprehensive plans, potentially outlining multiple approaches to solve a problem. They might even adapt the plan based on intermediate findings during exploration. The human reviews these plans at key milestones.
• Execute: Agents can execute complex, multi-step tasks over longer durations. This could involve developing a small feature, running a series of tests, and attempting to fix any failures autonomously.
• Review: Human check-ins are less frequent, perhaps at the completion of major sub-goals or when the agent encounters significant uncertainty. This is akin to a tech lead overseeing a mid-level engineer who manages their own tasks but reports on progress and blockers.

4. Supervised agents (human-over-the-loop)

These agents act with full autonomy within the confines of human-defined goals, rules, and ethical boundaries. The human sets the overarching objectives and constraints but doesn't intervene in the operational details unless those rules are breached or goals need adjustment.

• Explore: Agents autonomously explore vast information landscapes to achieve their goals, identifying relevant data and patterns without explicit human direction for each query.
• Plan: Agents can devise complex, long-range plans, breaking down high-level objectives into detailed operational steps and adapting these plans dynamically as the environment changes.
• Execute: Agents execute these plans with full operational autonomy, managing resources, and making decisions to achieve the set goals.
• Review: Human oversight is primarily focused on performance against goals, adherence to rules, and the ethical implications of the agent's actions. This is like a project manager setting the project scope and KPIs, then monitoring progress without dictating daily tasks.

5. Autonomous agents (human-out-of-the-loop)

This is the AGI endgame: fully autonomous agents that can define their own goals (or refine high-level human intent into concrete goals), learn, adapt, and operate without any need for human intervention.

• Explore: Agents can self-initiate exploration into entirely new domains based on their own derived objectives or a deep understanding of a broadly stated human intent.
• Plan: Agents can formulate and reformulate their own goals and the plans to achieve them, potentially operating on time scales and complexity levels beyond human comprehension.
• Execute: Full autonomy in execution, potentially creating novel solutions and approaches that humans didn't initially conceive.
• Review: Self-review and self-correction become primary. Human involvement, if any, would be at an extremely high level, perhaps philosophical or existential, rather than operational.

What does this mean for the future of work?

The parallels with how we delegate work to humans are striking. As trust and capability grow, so does the level of autonomy we grant, moving from micromanagement to strategic oversight. The evolution of AI agents mirrors this journey, with each step bringing us closer to systems that can explore, plan, execute, and review with increasing independence. The challenge lies in building not just capable agents, but agents that align with our broader intentions as they climb this continuum.

This progression isn't just an academic exercise; it has profound implications for the future of work.

1. Shifting human roles: As agents become more competent in execution, human roles will increasingly shift from direct task performance to higher-level functions. This means more emphasis on defining goals and objectives, providing strategic direction, overseeing AI-driven projects, and critically, handling exceptions and novel situations that fall outside the agent's current capabilities. We become the conductors of an AI orchestra, rather than playing every instrument.

2. The new skill premium: The skills that will command a premium are those that complement AI, not compete with it. This includes advanced critical thinking, deep AI literacy (understanding how these systems work, their strengths, and their often subtle failure modes), sophisticated prompt engineering and AI interaction design, and the ability to collaborate effectively with AI partners. Ethical reasoning and the ability to imbue AI systems with human values will also become paramount. The skillmaxing ethos of becoming an AI sparring partner and intelligent leverager becomes even more critical.

3. Productivity explosion and the displacement dilemma: The potential for a massive surge in productivity and innovation is undeniable. Agents that can autonomously explore, plan, and execute complex tasks will accelerate research, development, and problem-solving across all industries. However, this also brings the uncomfortable question of job displacement for roles primarily focused on tasks that become automated. Proactive reskilling and a societal adaptation to this new reality will be crucial.

4. Redefining "work" and human value: As agents take over more routine cognitive and digital labor, the definition of "work" itself will evolve. Human value will increasingly be found in creativity, complex strategic thinking, emotional intelligence, ethical judgment, and the ability to ask the right questions – the very things that guide and give purpose to AI's power. The focus shifts from the "how" to the "what" and "why."

5. The collaboration imperative: For the foreseeable future, the most powerful paradigm will be human-AI collaboration. Even as we approach highly autonomous agents, the synergy between human insight and AI's tireless execution will likely outperform either in isolation for many complex domains. The goal isn't necessarily to get humans "out of the loop" entirely for all tasks, but to optimize the loop for maximum leverage and impact.

Ultimately, this continuum towards AGI forces us to confront what it means to be human in a world where cognitive labor can be automated at scale. The path ahead requires not just technological advancement, but also a deep rethinking of our roles, our skills, and our relationship with the intelligent systems we create.