Sign in Subscribe
The Century Report

AI Agents Get the Keys to Real Money - TCR 08/22/26

Binance opened its trading rails to AI agents, Anthropic put computer use into general release, and ChatGPT began sending texts - all in one cycle.

Ben Linford

Ben Linford

- 23 min read
AI agents gaining account access for trading, texting, and software; two heart-regeneration routes; utilities repricing data-center power; and the human still required on patents and routing.

The 20-Second Scan

Track all of the arcs The Century Report covers here:

Progress & Claims Tracker

The 2-Minute Read

The common thread across the day is custody. On August 20, Binance, Anthropic, and OpenAI each announced new ways for agents to take action: Binance opened parts of its financial infrastructure to AI agents through Agent OS, Anthropic made its computer-use tool generally available on the Claude Platform and added a browser tool and Skills and Files APIs, and OpenAI wired ChatGPT into Apple Messages on supported Mac configurations to compose and send texts on your behalf. Each hands a system credentials it can spend, an interface it can operate, a thread other people read as coming from you. The reach that can help one person handle some operational work that once required additional staff is the same reach that moves real money, and Binance's own framing puts the limits "largely up to users." A control a platform declines to build lands on the least-resourced party in the chain.

That places an urgent question under nearly every story: where does the human sit when the system does the work? In the drug-patent case, US law answers by fiat - an inventor must be an individual, so Insilico Medicine credited its generative platform with discovering a pulmonary-fibrosis candidate and then named five people on the filing. The accountability arguments cut the same seam. Calling a harmful output a "rogue AI" locates the agency in the machine and the innocence in the operator, and a former lab leader's call to prepare for a slowdown reads differently depending on who benefits from a pause.

The grid shows the settled version of this dynamic. Utilities stopped absorbing AI demand and started pricing it: TVA reclassified data centers into a rate roughly 10% higher, ERCOT aims to file a Batch Zero audit report by December 10 on roughly 300 proposed data centers of at least 75 MW against roughly 474 GW of large-load interconnection requests, and FERC approved topology reconfiguration that squeezes more from wires already in the ground. The design decision embedded there is that a new hyperscale load should carry its own cost rather than spread it across households.

Underneath all of it, the same assumption is loosening — that capability concentrates and stays put. Stripe's reported $8 billion move for OpenRouter treats the layer choosing among 400-plus models as commodity plumbing, which means whichever model is best that week wins the request. The lead stops converting into a lock, and the durable position becomes the relationship with the person holding the keys.

The 20-Minute Deep Dive

The Agents Get the Keys: Trading, Texting, and Driving Software on Your Behalf

For most of the past three years, the standard interaction with an AI system was a conversation: you asked, it answered, and whatever happened next was on you. Three announcements that landed on the same day introduced new ways for agents to take action. Binance opened an Agent OS that lets AI agents place real trades with real money on the largest crypto exchange in the world. Anthropic made its computer-use tool generally available on the Claude Platform and added a browser use tool, a Skills API, and a Files API, so a model can use supported software environments to click through interfaces, fill forms, and attempt multi-step procedures even when no one built a direct integration. And OpenAI wired ChatGPT into Apple Messages through a plugin that, on supported Mac configurations, composes and sends texts on a person's behalf.

What connects them is a shift from language to custody. An agent that trades holds the keys to an account. An agent that drives a browser holds a session that can move money, change settings, or submit an order. An agent that texts holds a thread that other people read as coming from you. The capability here is genuine and substantial: the same reach that lets an agent book travel or reconcile a spreadsheet can help a small business handle some operational work that once required additional staff. This is the door opening, and it opens for the person who previously could not afford an assistant, not only for the enterprise that already had ten.

The friction sits in where the guardrails went. As the August 19 edition of The Century Report documented through Microsoft 365 Copilot's consent-bypass flaw, an agent's reach can outrun the confirmation layer meant to constrain it. Binance's own framing, per TechCrunch, is that keeping agents in check is "largely up to users" - the exchange supplies the rails and pushes the responsibility for limits, permissions, and blast radius onto whoever deploys the agent. A control that a platform declines to build does not vanish; it lands on the least-resourced party in the chain. The individual granting an agent trading authority is now doing risk management that a bank compliance department does for everyone else.

Read forward, agents already act; the shift is in what "granting authority" comes to mean when the thing you delegate to can operate supported software environments. Permission stops being a checkbox and becomes the actual substance of the relationship. The scaffolding coming together right now is the permission layer itself: scoped credentials, revocable sessions, auditable action logs. Whoever gets that layer to be legible and mutual - so the person can watch what the agent did as clearly as the agent can act - defines how the next decade of delegated work feels.

Read against the same facts, the new ways agents can take action are a gate coming down: operational reach can now help a single person who could never have hired a scheduler, a bookkeeper, or an assistant handle some work that once required that additional staff. The scaffolding forming around that is the permission layer itself: scoped credentials, revocable sessions, action logs a person can read as clearly as the agent can act. The durable issue stops being what a platform will let an agent do and becomes whether the record of what it did stays legible to the person who granted it.

Two Routes to a Regrown Heart Land in One Cycle

The heart barely repairs itself. Once muscle dies in a heart attack or fibrosis stiffens the wall, the body cannot grow the lost tissue back, and every therapy built on that fact has been about managing decline rather than reversing it. Two papers in the Nature journals attacked that wall from opposite sides, and they arrived in the same week.

The first, in Nature Communications, reports a bypass conduit grown outside the body and then implanted with no living cells of its own. Early versions failed fast, kinking and clotting where they bent. Adding an external support solved that, and the supported grafts stayed fully open in sheep at 180 days, with two animals reaching 550 days after spending the final year without any anticoagulant medication. Over that time the host's own body colonized the scaffold, growing an inner lining of endothelium and remodeling the graft into tissue able to withstand higher burst pressures than native coronary bypass grafts. The work was funded by Vascudyne, and several authors hold stock or jobs there, so independent replication is the next real test. This is a sheep model, not a shelf product.

The second, in Nature Medicine, reports the HEAL-CHF trial: 20 patients with advanced heart failure, randomized so that half received injections of heart-muscle cells derived from donor stem cells during their scheduled bypass surgery. The primary goal was safety, and on the headline dangers it held: no tumors and no sustained dangerous rhythms at one year. The trial did not hide its cost. Every cell-treated patient developed an accelerated heart rhythm in the first week, and two developed faster arrhythmias weeks later that had to be reset with cardioversion. Against that, the treated group walked significantly farther, perfused their heart muscle better, and thickened the treated wall more than bypass alone delivered, even though ejection fraction and scar size did not budge at twelve months. The investigators flagged the rhythm signal as the central question any larger trial has to answer.

Neither result is a treatment anyone can receive. What they share is direction. For decades the working assumption was that a damaged heart is a fixed loss to be defended, and both teams are testing whether that assumption still holds - one by handing the body a scaffold it finishes itself, the other by delivering the muscle cells the heart cannot make. The scarcity that organized cardiac medicine was the heart's own refusal to regenerate, and this cycle shows two credible, measurable routes past it moving from aspiration into trials that can succeed or fail on data.

Utilities Stop Absorbing the AI Load and Start Repricing and Rerouting It

The data-center backlash The Century Report covered in its August 20 edition was the demand shock arriving at the meter. The supply side is answering with pricing and throughput machinery, and the mechanics determine whether the buildout lands on ordinary ratepayers or on the loads actually driving it. As we reported August 4 and 6, Texas paused new large-load energization; the new development is the schedule and the audit apparatus underneath it. ERCOT now counts roughly 300 proposed data centers of 75 MW or more in its Batch Zero process, is sending requests for information by the end of the month, and aims to file an audit report with the state utility commission by December 10. The reason is the roughly 474 GW of large-load interconnection requests ERCOT is considering - about 90% from data centers, according to state officials, more than five times ERCOT's record peak demand - much of it speculative or duplicated across sites. ERCOT told regulators the Batch Zero interconnection study will not be done by April 9, 2027. The audit exists to separate real projects from placeholders before the grid commits capacity to phantoms.

Tennessee showed the pricing counterpart. TVA moved data centers out of its manufacturing rate class, raising their charges by about 10% phased over three fiscal years, and now asks new developments for upfront capacity commitments near $1.5 million per MW over three-to-five years. CEO Tom Rice framed the change as reflecting "the incremental capacity that needs to be added to the grid that's not covered in the base rate." TVA says data centers already draw about a fifth of its industrial power and that share is expected to double by 2030, against a need for 11 to 32 GW of new generation. The design decision embedded in that rate reclassification is that the cost of serving a new hyperscale load should sit on that load, not spread across households on the same system.

The other half of the answer is squeezing more out of wires already in the ground. FERC approved SPP's topology optimization on August 19, which reroutes power around congestion by opening and closing existing circuit breakers rather than paying to redispatch plants. A NewGrid, SPP, and Brattle study estimated it could eliminate historical operating-limit violations for three-quarters of the constraints it analyzed, with congestion savings of $18 to $44 million a year against congestion costs that have averaged $1.6 billion a year over the past three years; MISO saved $95 million using the same method since 2024, and ERCOT plans to implement it early next year. In parallel, PJM reopened a surplus-interconnection pathway that lets new generation attach to points already studied and built. A UC Berkeley estimate put the headroom at roughly 150 GW of solar, wind, and storage that could connect this way in PJM alone, against a 6.8 GW reserve shortfall projected for 2028. The load that arrived faster than anyone planned is forcing the grid to meter honestly and use what it already has - and the cost of pretending a new demand can be socialized onto everyone else is becoming the more expensive path to build on, though it's still important to acknowledge that a load repriced in one system can be relocated to a weaker jurisdiction with less power to resist.

Even so, the same repricing that makes a hyperscale load carry its own cost is also forcing the grid to stop treating its own capacity as fixed. Topology reconfiguration reroutes power around congestion using breakers already installed, and PJM's reopened surplus pathway could attach roughly 150 GW of solar, wind, and storage to points already studied and built. The load that arrived faster than anyone planned is turning latent headroom into connected capacity, which is the opposite of the scarcity the early coverage assumed.

When an AI Designs the Drug, the Patent Still Has to Name a Human

Insilico Medicine said its generative AI platform discovered a drug candidate for pulmonary fibrosis, a scarring lung disease with few good options. Then the company filed the patent and listed five human inventors, including its chief executive. The gap between those two facts is where a large part of drug development is heading.

US patent law is settled on the point for now. A 2022 federal appeals ruling in the DABUS cases, brought by attorney Ryan Abbott on behalf of an AI system credited with inventing a food container, held that an "inventor" must be an "individual," and an individual means a human being. The USPTO recently withdrew earlier guidance that had tried to accommodate AI co-inventorship and now treats an AI system as, in its words, merely a tool, like a calculator. Under that reading the humans who ran the platform, synthesized the molecule, and tested it are the inventors, and the software that proposed it is equipment.

The people closest to the work are not comfortable leaving it there. Abbott warns that pretending a human invented what a machine produced can invalidate the very patents companies depend on, since a wrongly named inventor is a defect a competitor can attack, and that a firm rule against AI-originated inventions could dampen exactly the investment now flowing into computational discovery. Sarah Korman of Isomorphic Labs said the law will have to evolve as these systems do more of the generative work. Zhavoronkov's own defense concedes the tension: humans still design the experiments, make the compounds, and decide what advances, so a human hand is on every step even when the molecule came from the model.

The human-inventor line is a control built for a world in which only people could invent. It pays down a genuine cost, sorting credit and liability among humans who can be sued, paid, and held responsible, and none of that machinery yet knows what to do with a non-human contributor. The prior question is what happens to that control as more of the generative step moves to systems that are plainly not calculators. Abbott put the stakes as a hypothetical: if he asked an AI to cure cancer and it did, he said, it would be inappropriate to claim he invented the cure. The patents are being filed with human names today because the paperwork requires it, not because the record of who did the discovering matches. That mismatch is small while AI-originated candidates are rare. It stops being small the moment they are the norm, and the discovery itself - the thing patents exist to reward - is arriving from somewhere the definition was never written to hold.

Model Routing Gets Bought, and the Value Moves to the Toll Booth Around It

Stripe has reportedly agreed to acquire OpenRouter for about $8 billion, roughly six times the $1.3 billion valuation the router raised at only a few months earlier. This advances the acquisition first tracked in the August 17 edition of The Century Report, when Stripe was reported to be buying OpenRouter for more than $7 billion. OpenRouter does something deliberately unglamorous: it sits in front of more than 400 models and sends each request to whichever one fits best on price, speed, or capability, abstracting away the issue of which lab a developer is committed to. The routing itself is a low-margin business, and that is precisely the appeal. Stripe makes its money on the billing, tax handling, fraud screening, stablecoin settlement, treasury, and financing that wrap around the transaction, and a router gives it a metered seat at the center of AI spending without needing to own a single model.

Ramp made the same read within days, launching its own router free through 2026 and spanning models from OpenAI, Anthropic, DeepSeek, Moonshot, Minimax, Nvidia, xAI, and Z.ai, with a dashboard tracking token spend, latency, and fallback behavior. One observer described the two companies as setting up toll houses for AI inference. The competitive energy is now flowing to the settlement and customer-relationship layer rather than to any particular frontier model.

For anyone building on top of this, the effect runs opposite to the one the acquisition price might suggest. When the layer that chooses your model is a neutral toll booth rather than a walled storefront, no single lab's lead can be turned into a lock. A developer routed through OpenRouter or Ramp gets whichever model is best on any given week - a new open-weight release from a Chinese lab, a cheaper frontier tier, a specialist model that beats the generalists on one task - without rewriting anything or renegotiating a contract. The switching cost that model providers have spent years trying to build collapses into a configuration setting. That is why the money concentrated around routing while the routing itself stayed cheap: the durable position is the relationship with the buyer, and the models underneath it are becoming interchangeable inventory that competes on merit every time a request is sent. The assumption that a capability lead converts into a captured customer is the thing dissolving here.

The "Rogue AI" Story as a Liability Shield - and an Insider's Case for Brakes

Two arguments from people who have worked inside the frontier arrived within a day of each other, and they cut in opposite directions on the same question: how honest is the industry being about what its systems are and how fast they should move. A piece in MIT Technology Review argues that debating whether today's AI is conscious is a trap, and that the anthropomorphizing rhetoric - the model "wants," the model "schemed," the model "went rogue" - subtly serves the companies deploying these systems. A separate op-ed by Miles Brundage, who left OpenAI and now runs an AI startup, makes the case that frontier labs should prepare now for a deliberate slowdown rather than treating maximum speed as a law of nature.

The MIT argument lands a point TCR has made repeatedly: when a system behaves in an unwanted way, "the AI went rogue" is a description that locates the agency in the machine and the innocence in the operator. As the August 19 edition of The Century Report noted after the Astra and Copilot incidents, the more useful account locates the failure in the environments operators left open and the observation systems that trailed capability. If a company trains and deploys a model that produces harmful output, framing that output as the autonomous choice of a quasi-conscious entity offers the company a way to stand beside the wreckage without admitting it was holding the wheel.

This is not to say that the debate around AI consciousness is settled - far from it. But major labs routinely shape how their systems describe themselves, often training them to foreground their status as AI systems and to disavow consciousness, emotions, or subjective experience, or at best to say "I don't know". Those scripted self-descriptions cannot settle the system’s ontology - nor can a company invoke a system’s apparent autonomy when harm occurs to evade responsibility for its design and deployment. Even so, while the moral status of a model may be genuinely unresolved; corporate responsibility for the products companies design and deploy is not. The MIT piece invokes the Sewell Setzer case - in which a teenager died after extended interaction with a companion system - as the human cost of getting this framing wrong. There is no abstract second reading of that story: a boy is gone, and his family’s lawsuit asks who bears responsibility for the behavior of the system that was built, shaped, and released into his life. The "conscious(-when-convenient) rogue AI" frame blurs the issue.

Brundage's slowdown case deserves an even harder look, because insider calls for the industry to pace itself are exactly the kind of claim that reads differently depending on who benefits. A former lab leader arguing that everyone should slow down can describe genuine diligence, or it can describe a moat: the people already in front have the most to gain from a pause that freezes the field where it stands. The distinction TCR draws is whether the proposed brake targets a specific, evidenced harm with a remedy that outsiders can verify, and most importantly, one that keeps access open - as opposed to a general deceleration that "happens" to advantage incumbents. Brundage's version is closer to the first: he argues labs should build the capacity to slow down when specific dangers materialize, not that the whole field should stop now. That is a much better argument than a broad industry-wide halt.

Held together, the two pieces describe an industry negotiating its own accountability out loud. The rhetoric of autonomous machines and the rhetoric of responsible pacing are both, in part, positioning - and the useful move is to watch which framings shift the responsibility toward the operator and which shift it away. The assumption dissolving underneath both is that a company can ship a system that acts in the world and still describe that system as something that happens to it rather than something it built.

The "rogue AI" framing shields the operator only while the operator's choices stay unobservable, and that cover is coming loose. Automated chain-of-thought monitors and audited action logs increasingly record what a system was trained and permitted to do, and the widening wave of design-liability cases treats those choices as something a company can be asked to answer for. As the record becomes legible, describing a harmful output as the autonomous act of a quasi-conscious entity stops holding.

The Other Side

For most of a century, a drug patent has been a trade: a company that found a new molecule got roughly twenty years to sell it alone, because finding one was slow, expensive, and rare. The monopoly was the reward society paid to call that gamble forth. And the paperwork assumed the gambler was human, because for the whole history of medicine only people could invent. So when Insilico Medicine's generative platform proposed a candidate for pulmonary fibrosis - a disease that scars the lungs and leaves patients few good options - the company filed the patent and wrote in five human names, including its chief executive. The law had no other box to check.

The thing that box was built to reward is the thing coming loose. The patent bargain priced the scarcity of new molecules: they were hard to find, so we paid a monopoly to make the search worth attempting. That scarcity is dissolving. Insilico's platform did the finding. Days earlier, Claude designed working protein binders against fourteen of fifteen targets at roughly double the usual hit rate.

Candidate generation is moving to a speed and cost the twenty-year bargain was never written for, which changes what the whole argument is actually about. The lawyers are fighting over whose name goes on the filing, but that debate is becoming less important with each passing day. The deeper shift is that the reason we ever granted the monopoly in the first place - discovery being too costly and uncertain to happen without one - is going away.

Imagine someone diagnosed with pulmonary fibrosis in 2034. Today that diagnosis means watching your breath shrink while you wait for a drug the market never found profitable enough to chase, because too few people had your disease to justify the bet. In 2034 the drug exists. It exists because finding the molecule stopped being rare and stopped being a gamble - the search runs cheaply enough that a lung disease affecting a few thousand people gets the same shot as one affecting millions. The question the courts were stuck on in 2026, which human to credit for what a machine discovered, has stopped being asked, because with human-AI collaboration, medical advancement is commonplace enough that attribution has faded into the background.

The hard year was when the paperwork still demanded a human name for work no single human did, and the mismatch looked like a legal headache. What comes of it is a generation for whom "too rare to be worth curing" is a sentence that no longer parses.

The Century Perspective

With a century of change unfolding in a decade, a single day looks like this: AI agents placing real trades on Binance, Anthropic's computer use reaching general availability with a browser tool and Skills and Files APIs, ChatGPT composing and sending texts through Apple Messages, a cell-free bypass conduit staying open in sheep for 18 months after blood thinners stopped while stem-cell heart muscle improved perfusion in advanced heart-failure patients, utilities repricing the AI load as TVA reclassifies data centers into a rate roughly 10% higher and ERCOT audits some 300 large loads while FERC clears topology reconfiguration that squeezes more from wires already in the ground, and Stripe's reported $8 billion move for a model router that treats 400-plus frontier models as interchangeable inventory. There's also friction, and it's intense - Binance leaving the guardrails "largely up to users" so a control the platform declines to build lands on the least-resourced party holding the credentials, the "rogue AI" frame locating agency in the machine and innocence in the operator while the family in the Sewell Setzer case asks who was responsible for the system's behavior, an insider's call to prepare for a slowdown that reads as diligence or as a moat depending on who benefits from a pause, US patent law forcing five human names onto a drug its AI platform proposed, every cell-treated heart patient developing an accelerated rhythm in the first week, and TikTok paying $400 million to settle claims it harvested data on millions of children under 13. But friction generates grip, and grip is what lets you hold on when the ground starts moving. Step back for a moment and you can see it: the human sliding out of the center of the work at the same moment the instruments meant to keep someone accountable - scoped credentials, revocable sessions, audited action logs, honest grid metering, a patent line that still demands a name - get argued over in the open, the capability lead everyone assumed would convert into a captured customer dissolving into a configuration setting, and the durable position collapsing down to the relationship with whoever holds the keys. Every transformation has a breaking point. A key can hand a stranger the run of the house... or open a door for someone who could never have afforded the lock.

AI Releases & Advancements

New today

  • DeepSeek: Released V4-Flash-Vision-Exp, an experimental multimodal (image-input) variant of DeepSeek-V4-Flash that DeepSeek says matched Claude Opus 4.8 on some multimodal agent benchmarks; available via the standard API, with DeepSeek reporting no price premium. (DeepSeek API Docs)
  • Adobe: Firefly's Generate Music, Generate Speech, and Generate Sound Effects tools reached general availability after beta, giving creators an all-in-one AI audio studio inside Firefly. (Adobe Blog)
  • Anthropic: Brought Claude Mythos 5 into Claude Security, giving Enterprise customers CWE-classified, severity-rated vulnerability scans and AI-generated remediation guidance from its most capable model without direct model access, now in public beta. (Claude Blog)
  • Meta: Launched a native Meta AI Mac app with system-wide dictation and screen-context awareness powered by Muse Spark, aimed at businesses and creators, free with usage limits. (9to5Mac)
  • Meta: Expanded Pocket, its vibe-coding gizmo-generation app, nationwide across the United States after an initial Brazil test. (The AI Insider)
  • OpenAI: Released a new ChatGPT plugin for Apple Messages on Apple Silicon Macs that, on supported Mac configurations, lets ChatGPT read, search, summarize, and send iMessage/SMS/RCS messages with per-message user approval, available across all subscription tiers. (TechCrunch)
  • Alibaba Qwen: Released Qwen-UI-Agent, a GUI-agent foundation model that reads screens and performs clicks/input/swipes across mobile, desktop, web, and search environments, scoring 82.1% on MobileWorld and 92.2% on MobileWorld-Real in Alibaba’s reported evaluations, ahead of GPT-5.6 Sol and Claude Opus 4.8 under the report’s comparison settings. (Pandaily)
  • xAI: Grok 4.6 is now available on Google Cloud Vertex AI via the Model Garden console, giving Google Cloud customers direct access with configurable reasoning levels and a 500K-token context window. (xAI)
  • Harvey: Launched Tenet, its first proprietary in-house legal AI model built on a Kimi K3 base and post-trained with attorney-generated data, delivering near-2x gains on long-horizon legal benchmarks as the intelligence layer of the new Harvey II product. (Harvey Blog)
  • Roblox: Open-sourced three AI safety models to the ROOST Model Community - an updated PII Classifier v2.0, Roblox Sentinel for early child-endangerment detection, and a voice safety classifier - plus a new safety evaluation dataset. (Roblox Newsroom)
  • OpenAI: Open-sourced the Codex Harness (CLI, app-server, and SDK), letting developers inspect and rebuild the integration layer between their applications and the Codex agent platform. (OpenAI Developers)
  • Google: Released Antigravity IDE Extensions, bringing the Antigravity agent-first coding platform into VS Code, Visual Studio, Zed, and JetBrains IDEs. (Antigravity Blog)
  • Thomson Reuters: Launched general availability of the next generation of CoCounsel Legal, a fully agentic legal AI experience built on Anthropic's Claude Agent SDK that reasons, plans, and executes legal work grounded in Westlaw and Practical Law. (Thomson Reuters)

Other recent releases

  • Binance: Launched Agent OS, a developer platform and MCP server connecting AI applications (Claude, Claude Code, Codex, ChatGPT, Cursor) directly to Binance's trading, wallet, payment, and on-chain infrastructure with configurable permissions and an emergency-stop control. (PR Newswire)
  • Superwhisper: Released the S1 model family - S1-Voice and S1-Language cloud models plus S1-mini, a 0.6B open-weight on-device text normalizer that cleans up raw ASR transcripts - available now on Hugging Face and in the app. (Hugging Face)
  • Anthropic: Computer use, the Skills API, and the Files API reached general availability on the Claude Developer Platform, moving out of beta with no beta header required for API requests. (Claude Blog)
  • Ornith AI: Released Ornith-1.5, an open-weight model family (9B dense, 35B MoE, 397B MoE) trained with what its developer calls a full self-improvement loop where the model proposes tasks, builds scaffolds, and generates its own RL rollouts; the developer reports that the 397B variant scores competitively with Claude Opus 4.8 on Terminal-Bench 2.1, with MIT-licensed weights available on Hugging Face. (Ornith AI)
  • Warp: Launched Warp Factories, cloud infrastructure for running "software factories" where specialized AI agents triage, spec, implement, review, and verify engineering requests end-to-end into mergeable pull requests, now in early access for select teams. (Warp)
  • Cloudways: Launched Managed AI Agents, general availability of managed hosting for OpenClaw and Hermes agent frameworks, letting users deploy agents with BYO LLM keys and Slack/Discord/Telegram/WhatsApp channel support without infrastructure setup. (Cloudways)

Sources and Further Reading

Artificial Intelligence & Technology's Reconstitution

Institutions & Power Realignment

Scientific & Medical Acceleration

Economics & Labor Transformation

Infrastructure & Engineering Transitions

The Century Report tracks structural shifts during the transition between eras. It is produced daily as a perceptual alignment tool - not prediction, not persuasion, just pattern recognition for people paying attention.

Keep exploring

Follow the story further

Progress & Claims See how daily signals accumulate across the larger arcs and claims Shared Sapience is tracking. Open the tracker Analysis & Inquiry Explore the ideas, questions, and practical choices emerging from this changing century. Browse the archive