The Agent Watch
Briefing · Articles · Tools · About EN FR DE ES 中文 IT PT SV FI DA

Daily Briefing

29 July 2026 · 5 stories

🔥 Top story

01

Google adds invisible guards around its agents

Google added invisible checkpoints to its managed agent service. When an agent runs a tool (executing code, writing a file), a small test can block the action, fix it, or simply log it. The result: budgets can be capped, scheduled triggers can be set, and a dangerous command can be stopped before it runs. For a team using agents, it is like putting a guard at the door of every sensitive room instead of watching the camera afterwards.

02

The language that links agents to tools gets simpler

The Model Context Protocol, which acts as an interpreter between an agent and its tools, drops the mandatory hidden session. Each request now carries its own information, like a stamped envelope rather than a phone call. This lets traffic be split across machines, every call measured, and costs billed more precisely. For users, it opens the door to tighter cost and permission control without breaking existing setups.

03

OpenAI shows where agent work really goes

OpenAI has published the detailed story of eight scientific projects in which agents maintained, optimized, or rewrote code. The main result is not speed: it is that human work shifts from writing to verification. The researchers who got the best results spent their time defining tests, comparing outputs against an external reference, and handling maintenance. An agent that produces fast without verification simply moves time loss from upstream to downstream.

04

Anthropic defends open-weight models but demands tests

Anthropic takes a public position: downloadable-weight models should not be banned. The company calls for three targeted measures instead: limiting Chinese access to advanced chips, fighting industrial-scale copying, and mandating safety tests regardless of whether the model is open or closed. For any company integrating a new model, the signal is clear: a provider can disappear from the market, but a provenance file and a security test remain useful for years.

05

Tines sells a control room for enterprise agents

Tines has launched 3B, a platform for cataloguing, running, and monitoring the software, automations, and agents built by teams. Each run happens in an isolated space that disappears after use, like a meeting room torn down after the meeting. Credentials are masked and every action is logged. The product responds to a simple phenomenon: many IT departments discover agents they did not know existed. This line confirms that agent governance is becoming its own market.

📡 To watch

Hush Security raises $30M for agent identities

Akamai is investing in a platform that replaces permanent passwords with temporary, revocable rights.

Claude Mythos finds weaknesses in cryptography

Algorithms in use today are not compromised, but human proof remains essential before any publication.

StateAct wants to read the program, not the screen

A submitted paper proposes acting directly on files and page structure rather than on the screen capture.

Check your MCP servers before migrating

Legacy session, logging, and transport features still have at least twelve months before final retirement.

📊 Trend

Agents are gaining real decision-making power, and the industry response is becoming clearer: invisible checkpoints on the vendor side, simpler protocols on the infrastructure side, and mandatory tests on the policy side. The rule that keeps coming up everywhere: an agent without a trace and a permission cap is an experiment that ends badly, and the best products become those that make verification unavoidable rather than optional.