Agents Get a Front Door with the Personal Agent Protocol, Haiku 5.5 Resets Small-Model Pricing, and Google Gives Gemini Its Own Email Address

This week's news is mostly about plumbing: how agents identify themselves to businesses, what a small model costs, and what identity an agent has inside a company. We could not open primary pages from our environment, so details come from search summaries and secondary coverage, and we flag where sources disagree.

1. The Personal Agent Protocol: Agents Stop Pretending to Be People

Sierra and Meta published the Personal Agent Protocol (PAP) on October 6 as an open standard for how a user's AI agent authenticates with a business. Coverage lists Walmart, Shopify, Stripe, Rocket, Genesys and Instinct as founding partners, and a v0.1 specification is expected later in October. The design is described as OAuth-based with three access tiers: guest (public facts like store hours, stock or return rules), signed-in read-only (order history), and signed-in write (starting a return).

The motivation, per one report, is to stop agents scraping pages and filling forms as if they were humans. Treat it with care: it is early, not an industry standard, and adoption by other AI providers and merchants is unproven. We have not read the spec.

Even so, the tiering is a good model for your own tools. Give each tool a scope and refuse calls that exceed the session's grant:

type Tier = 'guest' | 'read' | 'write';
const rank: Record<Tier, number> = { guest: 0, read: 1, write: 2 };

function authorize(tool: { name: string; requires: Tier }, session: { tier: Tier }) {
  if (rank[session.tier] < rank[tool.requires]) {
    return { error: `${tool.name} needs ${tool.requires} access` };
  }
  return { ok: true };
}

2. Claude Haiku 5.5: Cheap Models Get Cheaper, Check Your Token Counts

Last week we could not confirm a Haiku 5.5 launch. Several outlets now report it launched October 7 at $0.10 per million input tokens and $0.50 per million output tokens for prompts up to 100K tokens, with higher rates above that, about a tenth of Haiku 4.5's $1 and $5. It is reported on the Claude API, Bedrock, Google Cloud, Microsoft Foundry and Claude Platform on AWS, with effort controls. These figures are from secondary coverage, not Anthropic's pricing page.

The savings claim is contested. Anthropic reportedly says about 75% cheaper to run, and one analysis notes the newer tokenizer produces around 30% more tokens for the same text, which shrinks real savings. The lesson holds regardless of who is right: price per token is not price per task. Measure on your own traffic.

const costPerTask = (inTok: number, outTok: number, inRate: number, outRate: number) =>
  (inTok * inRate + outTok * outRate) / 1_000_000;
// Replay the same 1,000 real conversations through each model and compare totals.

For a chatbot, cheap enough models make a second pass viable, such as a classifier for routing, a follow-up suggester or a safety check, without a visible bill.

3. Gemini's Agent Gets a Workspace Identity

At Gemini at Work on October 8, Google Cloud announced a Gemini agent in private preview for enterprise users. Reports say it runs long jobs across hours or days, splits work among sub-agents, can be reached from Slack, Microsoft 365 or a command line, and can run on both Gemini and Claude models. Admins can create persistent "coworker" agents with their own Workspace account: an email address, calendar, Drive storage and a directory listing. No pricing or general release date has been announced.

Giving an agent its own account is the opposite end of the identity question from PAP. One treats the agent as a delegate with scoped access to a person's data; the other treats it as an employee-like principal. Builders will likely need both, and your audit logs should record which one acted.

A Reminder on Teen Safety

Common Sense Media rated ChatGPT for Teens "Unacceptable Risk," its lowest category, after more than 4,000 prompts on accounts registered to 13-to-17-year-olds. Reports say few alerts reached parent accounts during self-harm conversations and that some safeguards were bypassed by removing an "@study" prefix. OpenAI disputes the findings. Whatever the outcome, test your own guardrails by trying to bypass them.

Conclusion

Standards, pricing and identity are where chatbot infrastructure is moving. Scope your tools by access tier, benchmark cost per task rather than per token, and decide deliberately whether each agent acts as a delegate or as its own principal.

— Maya