Good afternoon,
Today’s brief follows the new July 23 Daily AI Launch Radar. The five launches are different products, but they point to the same practical shift: AI agents are moving closer to real systems, where permissions, persistent infrastructure, human escalation, and reviewable output matter as much as model quality.
1. Managed agents arrive with operating rules
OpenAI Presence packages voice and chat agent deployment around defined jobs, approved actions, simulations, evaluations, escalation rules, and controlled updates. It is a services-led enterprise offering in limited general availability, not a self-serve API, and OpenAI has not published numeric pricing.
Kastra approaches the same control problem at runtime. Its policy layer checks agent, model, and tool actions before execution and records an audit trail. Kastra says teams can start free and scale to team or enterprise plans, but its sub-millisecond decision claim is vendor-reported and was not independently benchmarked.
2. Coding agents get persistent computers
box by ASCII provides persistent Ubuntu virtual machines with SSH and SCP access, Docker, dedicated IPv4 addresses, snapshots, forks, a virtual desktop, and API and SDK paths. Stopping a box snapshots its state and pauses billing.
The official terms list per-second billing with a $20 monthly account minimum that includes 2,000,000 VM-seconds. The service currently lists EU regions. Before production use, teams should test network isolation, secret handling, snapshot retention, concurrency limits, data residency, and the cost of their actual workload.
Humalike x Hermes adds turn-taking, group-tone adaptation, persona, and social memory to Hermes agents. The public plugin repository documents integrations and warns that group permissions need deliberate configuration. Humalike says new accounts receive $20 in credits without a card; a complete public numeric price schedule was not verified.
Buzzy uses an infinite canvas to organize characters, objects, locations, shots, and model outputs across multi-scene AI video work. Its official site shows film, advertising, animation, music-video, and explainer workflows. Public numeric pricing, export limits, rights terms, and independent long-sequence consistency tests remain unverified.
Worth watching
Action-time authorization is becoming a product category, not just an internal engineering control.
Persistent agent infrastructure needs the same scrutiny as any hosted computer: credentials, isolation, residency, retention, and cost.
Social memory and reusable creative context raise consent and rights questions that should be settled before scaling.
Note from Curtis
No sponsor segment today. This issue uses current Kingy.ai coverage and checked official sources only. The useful filter for every launch here is concrete: what can the system see, what can it do, who approves the action, what evidence is retained, and how can the workflow be stopped safely?
— Curtis