$ today_in_10_seconds
Salesforce made Claude the default reasoning model across Agentforce and Slack, and shipped itself back into Claude as a 37-skill plugin. Meanwhile Google's A2A protocol moved in with MCP under one foundation. Two consolidations in ten days, both about who controls the layer agents actually run on.
$ lead_story
Salesforce and Anthropic announce Claudeforce
Announced August 26, Claudeforce runs in three directions at once: Salesforce inside Claude as a plugin, Claude inside Agentforce as a reasoning model, and Claude as the default model in Slack. Salesforce expects to spend roughly $300M on Anthropic tokens this year, on top of a stake in Anthropic now worth about $5bn. The plugin ships 37 prebuilt sales skills — meeting prep, deal-health reviews, pipeline updates — running against live CRM data without leaving the Claude window, with pilot access now and open beta in September. The interesting half isn't Claude landing in Agentforce; it's Salesforce shipping its governance layer outward, into somebody else's client.
// why_it_matters: The moat is moving from the model to the permissions, audit trail, and business logic wrapped around it.
$ also_worth_attention
The three layers of agentic AI security
Nutanix's Oscar Wahlberg argues that treating agent risk as one problem produces incomplete architectures, and splits defense across model, application, and infrastructure layers. The framing is right — the controls that catch a malicious prompt are not the controls that contain a compromised agent's blast radius. Worth reading with the byline in mind: this is sponsored, and the infrastructure layer is conveniently where Nutanix sells.
link →Agent2Agent joins the Agentic AI Foundation alongside MCP
A2A became a hosted project of the Linux Foundation's AAIF on August 17, putting agent-to-agent delegation and agent-to-tool access under one governance roof. AAIF has gone from under 40 members to more than 250 in eight months. The unglamorous caveat: LFX data for the year ending August 25 shows three organizations accounting for over half of A2A's recorded contributions. Neutral governance and neutral maintenance are different things.
link →Nobody can prove your agent was authorized
The AI AGENT Act (S. 5051, introduced July 21) defines a custodial user agent as one acting under transparent, limited, revocable authority, and directs NIST to develop standards for verifying delegation and keeping auditable records. What it doesn't require is a verifiable evidence chain spanning every system between the user's instruction and the final outcome. NIST's own February concept paper scopes the initial effort to agents inside organizations and defers agents crossing company boundaries — which is exactly where disputes happen.
link →
$ bigger_picture
The trend behind the headlines
Nothing shipped this week made agents smarter. Claudeforce is a distribution and governance deal. A2A joining AAIF is a stewardship change with no spec bump and no API break. The AI AGENT Act is an attempt to make delegation provable after the fact. The capability question got settled quietly sometime last year, and what's left is the boring, expensive part: identity, authorization, audit, and who gets to own the surface where agents touch real systems. If you're building, that's where your differentiation is now — not in the reasoning loop.
NIST NCCoE — Software and AI Agent Identity and Authorization
The draft concept paper works through how existing identity standards apply to agents that act autonomously and take downstream actions, and NCCoE is still collecting feedback on scope. It's not a library, but it's the clearest available map of the questions your security reviewer will ask in six months — and reading it now is cheaper than retrofitting.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime