Walking Points: Work, Explained on a Walk

Walking Points: Work, Explained on a Walk

By Walking PointsBusinessManagement
Download on the App Store

Walking Points: Work, Explained on a Walk episodes

  • Part 1: What an AI Agent Actually Is: From answering to acting

    Made for you, by AI, from the sources below.

    Updated in October 2026: METR's latest measurement is now Claude Mythos, at least 16 hours, published in May. Salesforce now reports 7 billion units of agent work to date.

    The first of five walks on AI agents at work: What actually separates an agent from a chatbot, why the capability curve crossed a real threshold in 2026 and not before, and three deployments with the numbers to back them, a coding agent that factored a 260-digit number, an enterprise platform past a billion dollars in revenue, and a customer-service assistant that had to publicly walk back its own hype. Closes with how to scope a first agent at work without repeating Klarna's mistake.

    Read the transcript

    In this episode

    • From chatbot to agent: what actually changed
    • Why 2026, and not last year
    • The deployments that made it real, and the one that overreached
    • What this changes for your work
      • Figures
      • 320 minutes: Claude Opus 4.5's 50%-reliability time horizon in METR's Time Horizon 1.1 release source
      • 88.6 days: The doubling period of METR's time-horizon metric since 2024 source
      • 4,900 GPU-days: Compute Devin used to factor RSA-260 autonomously over three weeks source
      • $900 million: Devin's annualised run-rate revenue at Cognition's September 2026 Series E source
      • $1.2 billion: Salesforce Agentforce's annual recurring revenue, up 205% year over year source
      • 3.8 billion: Agentic Work Units delivered through Agentforce to date, as of Q1 FY2027 source
      • 700 agents: Human customer-service capacity OpenAI estimated Klarna's AI assistant replaced source
      • 40 percent: The cost reduction of Claude Opus 5.5 versus Opus 5 per Anthropic's own announcement source

        • Sources: Deep dive

        • AI agent vs chatbot: the differences that matter in 2026 (DevRev)
          Source of the tools/memory/autonomy framing and the 'chatbots read, agents act' line used to open the episode.
        • Building Effective Agents (Anthropic)
          Origin of the agentic-loop framing (observe, reason, act, evaluate) used to define what makes something an agent.
        • Introducing OpenAI o1 (OpenAI · 2024-09-12)
          First of the technical unlocks credited for the 2026 capability threshold: a reasoning model that plans before acting.
        • 3.5 models and computer use (Anthropic · 2024-10-22)
          Second technical unlock: the first frontier model given general control of a computer screen rather than a bespoke API.
        • MCP joins the Agentic AI Foundation (Model Context Protocol / Linux Foundation · 2025-12-09)
          Third unlock and its maturation into vendor-neutral infrastructure, cited for the MCP origin story and server-count figures.
        • Time Horizon 1.1 (METR · 2026-01-29)
          Primary source for the time-horizon figures (Claude Opus 4.5, GPT-5, o3, GPT-4) and the doubling-period data anchoring 'why 2026, not last year'.
        • Introducing Claude Opus 5.5 (Anthropic · 2026-09-22)
          Most recent capability and cost jump, released the day before this research, used to show the curve is still compounding.
        • Factoring RSA-260 (Cognition · 2026-09-03)
          Primary account of Devin's three-week autonomous run, used as the episode's clearest example of sustained unsupervised agent work.
        • Do it all with Devin: Announcing our Series E (Cognition · 2026-09-08)
          Source for Devin's revenue growth, valuation and named enterprise customers, used as the software-engineering deployment example.
        • Salesforce Delivers Record First Quarter Fiscal 2027 Results (Salesforce · 2026-05-27)
          Source for Agentforce's ARR, growth rate and work-unit volume, used as the enterprise-platform deployment example.
        • Klarna AI assistant handles two-thirds of customer service chats in its first month (Klarna · 2024-02-27)
          Primary figure for the customer-service deployment's early results.
        • Klarna (OpenAI)
          Source for the '700 human agents' capacity-equivalent figure used to size Klarna's deployment.
        • Klarna AI Saved $40M on Support, Then Walked It Back (Twig)
          Source for Klarna's 2025 correction and the CEO's June 2026 'VIP thing' quote, used as the deployment that overreached.
        • The State of AI: Global Survey 2026 (McKinsey)
          Source for the adoption-versus-scaling gap cited in the closing section on starting a first agent at work.

          • Updated in October 2026

          Want a show like this about your own work? →

          16 min
        • Trailer: AI Agents at Work

          Made for you, by AI, from the sources below.

          In 2026 AI stopped only answering questions and started doing the work: Booking, coding, filing and deciding. Five parts on what AI agents actually do in workplaces today, where they fail, what they change for the people around them, and the rules that now apply. Researched and hosted by Mia and Adam, two AI voices, for anyone whose job is about to meet one.

          2 min

        About Walking Points: Work, Explained on a Walk

        From the publisher's feed

        Free seasons of Walking Points: Short arcs on one subject at work at a time, made for a walk and hosted by Mia and Adam, two AI voices. Made by AI from the sources in each episode's show notes, and…