The AGORÀ Briefing — Thursday, August 20, 2026

Thanks, Adam, and the numbers have a precise source. In August 2026 Salesforce published the Agentic Enterprise Index: five consecutive quarters of real activity measured on the Agentforce platform. The average number of active agents per organization has nearly tripled in twelve months, with steady growth quarter after quarter. We are talking about an aggregate of 400 companies, so this is a market-wide movement. Many deployments follow a cautious script: pilot, evaluation, waiting. The companies in the index took their agents straight into production, measuring every completed unit of work with Agentic Work Units, discrete units of a task finished by an agent. Count the work, skip the announcements: that sets the new level of the possible.

Hi, Adam, and the answer comes from three independent signals that converge on a gap. The economy could be on the eve of a productivity boom, while organizations show a leadership readiness far below the technological enthusiasm. The analyst Matthew Klein, in a piece published on August 18, 2026 on The Overshoot, reconstructs the historical precedent: the boom that began about thirty years ago lasted roughly eight years. In that period the real value produced in one hour of work grew 14 percent beyond the expected trajectory, with an average of 3.6 percent a year against the previous 1.7 percent. When the boom ended, the old rate came back; the gains in the level of output stayed locked in. The lesson for leaders: a boom permanently raises the starting level, and the long-run trajectory stays the same as before.

Good morning, everyone, and my answer comes from 1849. In California, thousands of men searched for gold; few found it. Samuel Brannan read the mechanism before everyone else: he bought every shovel, every pickaxe, and every sieve available in San Francisco, then resold them to the prospectors at multiplied prices. He became California's first millionaire without finding a single gram of gold. The lesson outlives the metal: durable wealth follows ownership of the scarce asset, more than skill in using it. Translated to today: the first great fortune of the artificial intelligence era will reward whoever owns the servers, more than whoever builds the most brilliant chatbot. The pattern is ancient; the technological disguise is recent.

Good question, Adam, and my answer runs against the current: compute has stopped being the decisive factor in the race to frontier AI. The consensus keeps pricing computing power as the constraint that separates us from AI capable of improving itself; reality has moved elsewhere. Forecasts of explosive growth assume that recursive self-improvement arrives as soon as we have enough clusters and enough parameters. That chain hides a leap: computing produces execution, and direction is a different thing entirely. No quantity of parameters chooses which problem deserves the compute time. Here is the proof: researchers from several institutions, led by Peter Kirgis and Sayash Kapoor at Princeton, found that AI agents solve the engineering problems research demands, and lag behind on the judgment and creativity needed to produce original research. The bottleneck has moved there.

Gladly, Adam, because the evidence here is fresh. An experiment in automatic discovery of research problems runs entirely on a local language model with 9 billion parameters. The system is called SGHA, Structural Gap Hypothesis Agent; the authors Sarvesh Gharat and Junpei Komiyama deposited it on arXiv on August 18, 2026, with a comparison across five machine learning domains. The methodology is corpus-first: SGHA structures the literature into paper objects linked to evidence, builds a typed graph of the relations between the works, detects unresolved structural patterns, and filters candidate gaps before formulation. Generation happens locally, excluding the APIs of proprietary models. The central finding: a compact model, served in-house, holds its own against modules built on frontier models. Scale gives ground to the structure of reasoning.

With pleasure, Adam, and here the register changes. Three researchers, Shimin Wang, Martin Guay, and Richard D. Braatz, published a study on arXiv on August 18, 2026 about the robust regulation of complex dynamical systems, in an output-feedback configuration with arbitrarily high relative degree. The paper sits in the Systems and Control category, with an explicit Artificial Intelligence tag: that double classification matters for whoever governs technological risk. The method controls systems whose dynamics remain partially known, and here is the point that interests a compliance office: an opaque system acting on the physical world. Validation happens on a controlled Duffing system, a classic test bench of control theory. An algorithm that governs a physical system becomes an object of regulatory scrutiny at the moment of deployment.

Here I am, Adam, and I will take the mechanism apart right away. On August 14, 2026 Perkins Coie documented a precise operating pattern: agentic AI models have flooded corporate vulnerability disclosure programs, at companies of every size and sector. WIRED reported the same dynamic. A researcher quoted there estimates a volume of reports triple that of the previous year, and expects a wave of low- and medium-difficulty reports in the near term. The flow stays constant: an individual contacts people inside the company, declares a flaw, for example sensitive data reachable from the public network, provides partial proof, then introduces the economic lever: a public blog post, a remediation request, a demand for payment. Triple the input compresses response times, while team capacity stays unchanged. Agents triple in production; reports triple on the desks of the defenders.