As AI started doing the work, I checked the permissions and the bill

Investment commentary in this article is market-analysis information, not a recommendation to buy or sell any asset.

As of the morning of August 7, 2026, Korea Standard Time

My first thought while reading today’s material was that asking whether AI gives a good answer is no longer enough. Once it creates documents, executes code, and begins touching orders and deployments, the benchmark needs a permissions table, a recovery procedure, and a bill beside it. It feels less like hiring one clever employee and more like meeting someone who asks for an access card, a corporate card, and the server-room key on the first day.

I did not repeat yesterday’s Muse Code, Cloud Run Sandbox, Foxconn AI-server, Circle, central-bank, gold, and KOSPI coverage. I also did not copy the supplied draft verbatim. Where an official page showed a different publication date, or where a product name and number could not yet be verified in a searchable primary source, I narrowed the claim. I chose links I can revisit over numbers that merely make a strong headline.

My main judgment today

The next stage of AI competition is about completed work, portable tools, controlled action, and investment returns. Models still matter, but the period in which a model alone could look like a complete product is ending quickly.

I am especially cautious about a smooth sentence that says higher use proves productivity and higher productivity justifies every infrastructure project. Each step needs its own evidence. Did the task finish? Did human repair fall? Did the system obey policy? Did the investment create enough cash flow to repay its financing? Those are separate checks.

1. ChatGPT use was deepening from questions toward deliverables

OpenAI Signals aggregates usage on individual ChatGPT plans and describes work-related use and adoption across countries. The date on the official page is June 30, not the August 6 first release described in the supplied draft. The research also uses a 0.1% sample of eligible individual accounts rather than every account. I therefore did not present it as a newly released August 6 dataset covering one billion users.

The direction still matters. At work, document creation, analysis, coding, and other deliverables are becoming more important than asking a standalone question. I see the same distinction in my own work: whether a file exists, a test passes, and a record remains for the next person matters much longer than whether an answer sounded impressive.

AI KPIs need to move with that reality. I would measure task-completion rate, human edits, time to completion, call cost, and recovery after failure—not accuracy alone. Message volume can prove that everyone is busy; it is not a receipt proving the job is done.

Sources: OpenAI — How ChatGPT adoption has expanded, OpenAI Signals report

2. Agent Plugins 1.0 is a promising package format, but still a working draft

Agent Plugins Specification 1.0.0 defines a package containing plugin.json, Agent Skills under skills/, and MCP server configuration in mcp.json. The goal is to let clients discover and run a Skill and its tools without rebuilding the package for a single agent platform.

The official document currently labels the specification a Working Draft. The supplied briefing described Google, Amazon, Microsoft, OpenAI, and Vercel as if they formed a settled standards consortium. The technical charter I could verify instead says governance seats belong to individuals, reserves no seats for companies, and prohibits one vendor from controlling a majority. I therefore focused on the format and its current status rather than a corporate roster.

This could become something like package management for agents. If it works, teams will stop wrapping the same capability separately for every IDE. The package, however, can carry executable components and MCP connections, not merely documentation. Signatures, provenance, authorization, secret injection, update history, and rollback are part of the product. A convenient supply chain remains a supply chain—and therefore an attack path.

Sources: Agent Plugins Specification 1.0.0, Agent Plugins technical charter

3. Kimi K3’s announcement and pause made deployment state visible

Moonshot AI’s open-weight Kimi K3 is offered by Fireworks AI at $3 per million input tokens, $0.30 for cached input, and $15 per million output tokens. GitHub announced that it was adding Kimi K3 to Copilot, then temporarily paused the rollout while responding to a GitHub Actions incident. At publication time, GitHub’s supported-model documentation still listed Kimi K2.7 Code, not K3.

That sequence shows how much the verb “supports” can hide. Announced, deployed to a cohort, enabled by an organization policy, and available in every client are different states. Even when the model is the same, hosting, caching, tool execution, and the IDE or cloud-agent path can change performance and the blast radius of an incident.

An enterprise needs replaceable routing and regression evaluation more than a permanent winner. I would run the same repository task across models and compare test pass rate, unrelated files changed, latency, cost, and policy violations. Choosing a new model is the fun meeting; building the rollback button is usually the valuable one.

Sources: Fireworks AI — Kimi K3, GitHub Copilot supported models

4. Agent security was moving from prompt filters toward action control

Policy in Amazon Bedrock AgentCore intercepts agent-to-tool traffic at the Gateway and evaluates calls with Cedar policies outside the agent code. Its important property is deterministic enforcement before a tool invocation and an auditable boundary independent of the model’s reasoning.

I could not independently verify the supplied draft’s new Dogwood language or its claimed sequence and cumulative-limit features in the AWS sources available while writing. I therefore did not publish that name as a confirmed launch. The AgentCore Policy capability I could verify already establishes the useful direction: do not rely on the model’s goodwill or a prompt for authorization. Enforce policy outside its execution path.

If an agent can pay or buy, a single-call permission and a workflow rule must be different controls. Permission to call an order API does not imply permission to place it a hundred times today. I want user-specific limits, sequence, cumulative exposure, reapproval conditions, and a kill switch in the policy layer. A well-written “please don’t” is less reassuring than a door that simply will not open.

Sources: AWS — Policy in Amazon Bedrock AgentCore, AWS — Why AgentCore Policy chose Cedar

5. The same AI pressured development services and created security demand

EPAM reduced its 2026 revenue-growth outlook from 4.0%–6.5% to 3.2%–4.2% and reported a 1.3% decline in software and high-tech client revenue. Gen Digital, which operates Norton and other security products, raised its annual outlook as demand for security and identity protection remained strong. Put together, the releases make “AI grows all of IT” look much too broad.

Code generation can reduce the labor required for repetitive implementation and maintenance, pressuring the price of traditional staff-based services. The same technology lowers the cost of producing phishing copy, malware variations, identity fraud, and automated attacks. Savings on one side can become defense spending on the other.

As a developer, I am reminded that implementation speed alone is a weak long-term shelter. Architecture, domain judgment, security, verification, incident ownership, and control over AI output become more valuable. As an investor, I would separate revenue AI removes from revenue AI creates instead of accepting one “AI exposure” line.

Sources: EPAM investor relations, Gen Digital investor relations

6. Terafab: I separated the ambition from the definitions behind its numbers

Reuters reported that SpaceX and Tesla would initially commit $16.8 billion to Terafab in Grimes County, Texas. The proposed complex would integrate logic and memory-chip manufacturing, packaging, and testing. Its long-term ambition is to support more than one terawatt of computing demand.

Earlier local public notices described $55 billion for initial phases and up to $119 billion with later expansion. The $16.8 billion figure may be a newly committed first tranche while the larger figures describe a multiphase total, but numbers with different scopes and dates should not be mixed as one project cost. I treated today’s figure only as the reported initial commitment.

Vertical integration can reduce supply constraints and shorten the loop between design and manufacturing. A semiconductor fab does not become competitive simply because it is expensive. Yield, process stability, equipment, power, water, skilled labor, and utilization still have to work. Software needs hardware abstraction for a world beyond one accelerator stack, while investors first need a reason the enormous plant will keep running.

Sources: Reuters — initial Terafab investment, Supply Chain Dive — Terafab plans and investment ranges

7. Tokenization depended more on legal ownership than on a coin price

The supplied briefing said Schroders had received Central Bank of Ireland approval for a tokenized share class of a U.S. dollar money-market fund. I could not sufficiently cross-check the approval details in the central bank’s public material and a Schroders primary release while writing. I therefore did not present the approval and product name as a confirmed core item.

The broader direction is verifiable. J.P. Morgan’s Kinexys provides infrastructure for asset managers to tokenize regulated money-market fund shares, and the Central Bank of Ireland has publicly discussed both the opportunity and regulatory foundation for tokenized finance. This is not permissionless trading of a public cryptoasset. It keeps the fund’s legal structure, KYC, custody, and redemption rules while connecting ownership records and settlement to distributed ledgers.

In an RWA system, the hardest work may be offchain legal ownership, investor allowlists, NAV, redemptions, and synchronization with the custodian—not the smart contract. I also resist assuming that the price of a chain token rises automatically with tokenized assets. Where infrastructure value turns into fees and cash flow remains a separate question.

Sources: J.P. Morgan — Tokenized Money Market Funds on Kinexys, Central Bank of Ireland — approach to tokenized finance

8. Productivity was welcome, but one quarter did not prove an AI revolution

U.S. nonfarm business labor productivity rose at a 1.4% annualized rate in the second quarter, while unit labor costs increased 1.3%. Better-than-expected productivity combined with softer labor-cost growth is encouraging for both corporate efficiency and inflation. Some economists said AI adoption may be beginning to contribute.

Possibility is not causality. Productivity also moves with the composition of output, hours worked, investment lags, and base effects. I want to see whether output per hour improves across several quarters and whether AI-heavy firms actually improve deployment cadence, incident counts, lead time, and output per employee.

On a development team, I would not use lines of generated code as a productivity measure. More code can mean more maintenance. Faster delivery, fewer incidents and repairs, and less customer waiting are more honest measurements.

Sources: U.S. Bureau of Labor Statistics — Productivity and Costs, Federal Reserve — AI, the economy, and the financial system

9. Oil reversed in a day, so I chose traffic over headlines

Renewed concern about restrictions in the Strait of Hormuz pushed Brent futures up 3.83% to $82.49 a barrel and WTI up 2.75% to $77.29, according to market reports. The move reversed the previous day’s decline on negotiation hopes. Spot and futures prices vary by timestamp, so I did not combine them as if they were one quote.

An AI data center does not directly burn crude, but an oil shock can travel through natural gas, electricity, logistics, construction materials, and long-term power contracts. A team can optimize token cost and still find the bill entering through another door if it ignores electricity and depreciation.

In a fast-moving geopolitical market, actual vessel traffic and supplied barrels matter more than alternating statements that talks are progressing or restrictions are under review. I will not extend today’s direction into a long trend before the physical flow changes.

Sources: U.S. Energy Information Administration — crude spot prices, U.S. Energy Information Administration — Strait of Hormuz

The judgments I am keeping today

1. Measure completed work, human repair, and recovery after failure—not AI usage alone.
2. Plugins and multimodel systems improve portability, but require provenance, authorization, supply-chain controls, and rollback.
3. When agents act in the world, enforce policy outside the agent code.
4. Judge AI infrastructure by yield, utilization, financing, and cash flow rather than announced scale.

My one-line conclusion is this: AI is taking on more work, but durable advantage will come from proof that the work finished, boundaries that stop the wrong action, and operations capable of earning back the investment.

Investment commentary in this article is market-analysis information, not a recommendation to buy or sell any asset. Every investment decision and its consequences remain the investor’s responsibility.