Daily Briefing
July 24, 2026 · 5 stories (site) · 7 stories (base)
On July 24, 2026, agents move into the infrastructure: HubSpot coordinates them on a single dashboard, AMD launches a full rack to host them, DevRev opens the phone channel with shared memory, YubiKey adds a human validation button, and telli funds multichannel customer relations.
🔥 Top story
01
HubSpot finally pulls every commercial agent onto a single dashboard
A sales or support team no longer has to juggle several robots that do not talk to each other. HubSpot is opening Agent Hub in beta, a central place to build, monitor and coordinate the sales, marketing and customer-service agents. They all work from the same customer history, which cuts down on contradictory actions. It is a bit like moving from a whiteboard covered with sticky notes to a single screen where everyone sees who is doing what. HubSpot adds Agent Builder, which lets a manager describe the work to automate in plain language, then link triggers, HubSpot data and third-party apps. A customer cited by HubSpot, Ignite Reading, says it went from fifteen to twenty minutes to a few seconds when compiling the calendar of every school district, saving more than 350 hours a year. Profitability will now depend on proving that coordination lifts sales or cuts service cost, beyond the comfort of better management.
02
AMD launches Helios: a full rack built to run AI agents at scale
For an AI agent to respond quickly to millions of people, it needs solid infrastructure. AMD is introducing Helios, a compute rack that brings together 72 MI455X accelerators, 18 EPYC processors and the ROCm software stack. Anthropic plans to deploy up to 2 gigawatts of it, and OpenAI is targeting a first rollout in the fourth quarter of 2026. In practice, a customer service operation that handles the volume of a large phone carrier could now rely on machines that are no longer only Nvidia. AMD claims up to 30% more text generated per dollar against the competing reference solution; that figure is supplied by AMD and still needs to be confirmed in production. The ROCm.ai platform is also meant to help Claude, Codex and Cursor better understand and tune AMD equipment. It is another step toward a real alternative for the teams running agents in production.
03
DevRev takes its agents from keyboard to phone call, with the same memory
A customer who calls about a problem should not have to start all over again because they switched channels. DevRev is extending its support agent to live phone calls, sharing the same company memory as the written exchanges: customer files, orders, tickets and documentation. The agent can resolve a case during the call or hand the full context to a human advisor. It is the difference between an answering machine that takes messages and a colleague who already knows the file. Voice AI is integrated into DevRev Computer, keeps the existing access rights, records and transcribes the calls, and handles several languages. The company claims 94.3% accuracy on its own internal test; that figure is a vendor measurement, not an independent validation. For companies, the appeal is mainly being able to turn this service on in a few hours, without changing the phone system.
04
YubiKey 5.8: a physical button to approve a sensitive AI agent action
When an AI agent prepares a transfer, signs a document, or changes a critical file, the last step is often missing: human confirmation. Yubico is extending its physical YubiKey key so that it can now let a person approve a specific action prepared by AI. It is like requiring a second signature on an important cheque: the machine proves to the system that someone actually pressed a physical button, something that cannot be faked remotely. Version 5.8 supports CTAP 2.3 and ships a preview of a WebAuthn extension for signing targeted actions. It is available on the YubiKey Bio, YubiKey 5 and Security Key ranges; the FIPS and CCN versions remain on 5.7.4. For companies that hand important rights to agents, this is a concrete guardrail that can be applied right away.
05
telli raises $15 million to sell its multichannel customer-relations agents
A Berlin startup named telli has raised $15 million to develop its customer-relations agents. These agents answer by phone, messaging, WhatsApp or email, qualify prospects and book appointments. The company says it already processes millions of conversations and is funding its internal agent Charlie, which orchestrates those channels. The round is led by redalpine, with Mutschler, Cherry Ventures, Y Combinator and several individual investors; cumulative funding now exceeds $18.5 million. This round confirms that investors are now backing agents applied to a measurable business function, rather than simple generalist assistants. The capital will be used for hiring and extending the multichannel platform, in a market where every saved call counts.
📡 To watch
AMD–Cerebras real-world performance still to be confirmed
The partnership promises up to five times more throughput, but the figure is based on an internal model. Availability, pricing and production gains are not published yet.
Chinese controls on model weights still in consultation
Consultations on export restrictions for Chinese models are reported, but no final text from the Chinese Ministry of Commerce is available at this stage.
Agent Hub profitability will depend on sales measurement
The HubSpot beta will have to prove it lifts sales or cuts service cost, beyond the coordination gain alone.
DevRev accuracy figures are vendor claims
The 94.3% accuracy claimed by DevRev on its internal test is not independent validation; real-world performance will depend on customer environments.
📊 Trend
Agents are becoming a full infrastructure layer, not just another gadget to plug in. HubSpot places them at the centre of the commercial workstation, AMD dedicates a whole compute rack to them, DevRev opens the phone channel to them, and YubiKey adds a human validation button. At the same time, the risk of a regulatory switch, whether it comes from Washington or Beijing, turns the choice of an engine into an architectural decision. The operational priority is now clear: measure business impact, keep a human for sensitive decisions, and plan a replacement engine.