Daily briefing
July 21, 2026 · 5 stories (site) · 8 stories (base)
On July 21, 2026, AI agents are becoming more independent and more exposed to supply risks: Kimi K3 opens the race for very large Chinese models, Washington considers limiting access to them, Inkling helps companies customise a model, Devin lowers the cost of coding work, and Natural prepares payments between agents.
🔥 Top stories
01
Kimi K3 puts a giant Chinese model within reach of coding agents
An AI agent that writes code or automates a task can now use a Chinese model of unprecedented size. Moonshot AI has released Kimi K3, a 2.8-trillion-parameter open model, available through a subscription or a service connection. It has beaten leading closed models in some coding tests and already works with several agent coding tools. Its announced price is below many US competitors. The model weights are promised before July 27, which could let teams run it on their own machines.
02
Washington considers cutting access to Chinese models — but nothing is signed yet
The White House is studying several ways to limit the use of Chinese AI models, Axios reports: sanctions, restrictions on public purchases and pressure on US companies hosting them. Kimi K3 is named as the trigger. For a company running an online agent, the risk is concrete: a model provider could become unavailable after a political decision, like a bank blocking a card. Important caveat: Axios reports internal discussions, not a published law or order. Do not change production on the basis of a rumour.
03
Inkling opens a new route: adapting a large model to your own business
Thinking Machines Lab, founded by Mira Murati, has released Inkling under a licence that broadly allows reuse. The goal is not to sell the best general model, but to provide a base each company can adapt to its documents and habits. It is like buying a solid house and changing the rooms, instead of renting the same flat as everyone else. Inkling handles text, images, audio and video, and can be customised with Tinker. For agents, this may mean more control over data and behaviour instead of relying only on a closed service.
04
Devin gets a faster, cheaper coding agent
Cognition, the company behind Devin, has announced SWE-1.7, the model powering its coding agent. It was improved from an open Chinese model and then trained by Cognition for software tasks. The company advertises about $1.97 per task in its main test and very fast answers through Cerebras chips. The difference matters: customers are not only paying for generated words, but for a result, like a software change ready for review. Agents are moving closer to professional work measured by what is actually finished.
05
Natural raises $30 million to let AI agents pay online
An agent that books a ticket, buys data or orders a service will eventually need to pay without asking its owner to click at every step. Natural has raised $30 million at an announced $200 million valuation to build payment infrastructure for agents. The project resembles a card network designed for software rather than people. Without a way to pay, an agent can only prepare actions; with a controlled budget, it can complete them. No direct company announcement was found, so the figures need confirmation before an investment decision.
📡 To watch
Alibaba promises open weights for Qwen 3.8 Max — with Qoder as the gateway
Alibaba presents a 2.4-trillion-parameter model available through its cloud offers and Qoder coding tools. Open weights are promised, but no date or licence has been published and the claimed performance remains internal. Watch it as a possible coding-agent alternative, not as a proven result.
Ten US health agencies will test OpenAI and Anthropic assistants
The PULSE programme should let ten public-health jurisdictions test AI assistants for health monitoring, translation and operations. OpenAI and Anthropic each provide ten enterprise licences. The direct CHAI announcement was not found, so the information remains to be confirmed.
Nous Research is reportedly valued at $500 million for its Hermes agents
The lab known for Hermes models reportedly raised $75 million at a $500 million valuation. Its ambition is to offer open models and agents for organisations and developers. The figure comes from a funding tracker, with no direct announcement found: treat it as a lead to verify, not a final fact.
Kimi K3 still has to pass the decisive test: publishing its weights
Moonshot promises to publish Kimi K3 weights before July 27. That will be the real test for teams wanting to run it themselves: hardware cost, installation and real behaviour will all need checking. The online release matters; the downloadable version will show whether the model can become a durable agent building block.
📊 Trend
The agent race is moving from chat to execution. Open Chinese models are putting pressure on prices and suppliers, while US labs are making customisation and coding agents easier. At the same time, an agent must be able to pay, switch models and work in a real profession. For Info-Sys, the priority is not an immediate model switch: keep several doors open, test costs on real tasks and never let one political or commercial decision stop an entire service.