$ today_in_10_seconds
Anthropic open-sourced a commerce-agent blueprint yesterday: working shopping and merchant agents, reference builds for retail, travel, telecom and ticketing, Shopify and Priceline already live. Then it drew a hard line: the agents can recommend and prepare actions, but they don't complete purchases. That line isn't caution for its own sake. It's the exact spot where the standards work isn't finished.
$ lead_story
Anthropic ships commerce agents with the wallet deliberately left out
Anthropic released a commerce-agent blueprint on September 2 - a public repo with two working agents and the harnesses, patterns and guardrails an engineering team needs to stand one up in days. The shopping agent handles catalogue search, multi-item carts, comparisons and service questions before handing off to checkout; the merchant agent covers sales analytics, inventory, pricing recommendations and campaign drafting with human approval required. It ships with reference implementations for retail, travel, telecom and ticketing, a Claude Code plugin, and deployment paths through the Claude API, Bedrock, Microsoft Foundry and Vertex AI. Shopify and Priceline are already running customer-facing agents, with Priceline having rebuilt its assistant Penny on Anthropic's models. Read the design decisions, not the logo wall. Human approval on campaign drafting and a hand-off at checkout are both admissions that delegated authority is the unsolved part. Anthropic's own numbers - carts up to 35% larger, shoppers 60% more likely to convert - are vendor-reported, so treat them as a hypothesis to test against your funnel, not a benchmark.
// why_it_matters: The forkable repo means the build is no longer the bottleneck - authorization is.
$ also_worth_attention
Anthropic pushes Claude deeper into agentic commerce as identity standards evolve
The commerce launch lands on identity groundwork Anthropic has been quietly laying since spring. Its updated privacy policy, effective July 8, introduced a Verification Data category covering government ID images, selfies and facial geometry templates, with checks running through Persona since April 14 and Yoti handling age-only pass/fail. Consumer tiers only - Team, Enterprise and Platform customers are exempt. The insight: an agent that spends money needs a verified principal behind it, and consumer KYC is the cheapest available proxy for one.
link →IETF charters DAWN as agent identity drafts pile up
The fortnight before September 2 brought a dense cluster of Internet-Drafts on agent authentication, authorization and delegation, converging on the newly chartered DAWN working group - which explicitly puts identity and trust management out of scope, leaving that to WIMSE and adjacent drafts binding short-lived, audience-restricted tokens to a human or organizational owner. DAWN's own remit is discovery: how a client learns what an agent is, how to reach it, which protocols it speaks and what it offers. The insight: vendor-specific agent auth has 12-18 months before the protocol layer catches up, so don't over-invest in bespoke delegation plumbing now.
link →Claude Fable 5.1 cuts cache reads 75%, the line item that dominates agent loops
Fable 5.1 holds Fable 5's $10/$50 per-million list price while dropping cache reads 75% to $0.25 per million, which Anthropic puts at roughly 25% cheaper for typical workloads and up to 45% for heavily agentic ones. That's half Opus 5's cache-read rate despite twice the base price - the result of a 0.025x cache multiplier where other models use 0.1x. The insight: repeated-context re-reads are what make a long-running commerce agent expensive, and this is the first pricing change aimed squarely at that shape of workload rather than at raw intelligence.
link →
$ bigger_picture
The trend behind the headlines
Three things shipped this week that only make sense together. Anthropic gave away the agent code, cut the price of the token category that agent loops actually burn, and stopped short of letting the agent pay. Meanwhile the IETF is scrambling to standardize the thing that would let it - provable delegation from a verified human to a scoped, short-lived agent credential. The capability gap in agentic commerce closed a while ago; what's left is an accountability gap, and nobody sells you a fork-and-deploy repo for that. If you're building here, the interesting engineering has moved from 'can the agent assemble the right cart' to 'who is liable when it assembles the wrong one.'
SPIFFE / SPIRE
Before your commerce agent goes near money, it needs an identity that isn't a long-lived API key in an env var. SPIRE issues short-lived, attested workload identities (X.509 or JWT) after verifying properties of the requesting process, so a rogue process on the same host can't just ask for another workload's credential. It's the substrate the IETF agent-auth drafts are profiling on top of, which makes it a reasonable bet ahead of the standards landing.
$ subscribe --free
Get the next edition in your inbox
// DAIly Agent covers agentic AI three times a week
// free, no spam, exit anytime