Decagon Builds a Gate for Customers' Own AI Agents, Convoso Ships a No-Code Voice Agent, and Four AI Labs Back One Robots.txt Opt-Out

Three stories from the past two weeks: a support vendor treating customers' own AI agents as a distinct kind of visitor, a contact-center platform turning past call recordings into a voice agent, and four rival labs agreeing on one way for sites to say no to AI training. Details come from vendor announcements and secondary coverage, so treat specifics as the reporters' own.

1. Your Next Customer May Be Someone Else's Agent

At its Dialogues event on October 1, Decagon announced four products, including a Personal Agent Gateway aimed at AI agents that contact a business on a customer's behalf. Coverage describes two parts: detection that flags likely personal agents in chat and voice, and a dedicated channel with its own operating procedures, authentication, limited access, and owner approval for sensitive actions. Decagon also introduced a consent and trust protocol called PACT and says the specification is available. Voice 3, announced alongside it, is described as running on a duplex architecture so the agent can listen while it talks. All performance claims are the vendor's own, and the Gateway details are less settled than Voice 3's.

The design question applies to anyone running a support bot: an automated caller is not a human typing slowly. It can retry endlessly, impersonate, or act without the owner's knowledge. Treat it as a separate tier with its own scopes.

type Caller = { kind: 'human' | 'personal_agent'; ownerVerified: boolean };

function allowedActions(c: Caller) {
  if (c.kind === 'human') return ['read', 'update_profile', 'refund_small'];
  // Agents get read access by default; anything sensitive needs the owner's approval.
  return c.ownerVerified ? ['read', 'request_refund'] : ['read'];
}

2. Voice Agents Trained From Your Own Call Recordings

Per Unite.AI, Convoso made its no-code Voice AI Agent generally available around November 2 to 3 (the two reports give slightly different dates). It qualifies leads and hands sales-ready prospects to human agents. Convoso's help center says setup involves uploading at least 200 past call recordings, then choosing a voice, an LLM provider and speech-to-text. Listed controls include do-not-call checks, state-level dialing restrictions, time-of-day rules and PII redaction. For now it works only with Convoso's own dialer. Some latency and training-data figures appeared only on a non-English page and we could not confirm them.

The pattern worth noting is compliance as a product surface. When a voice bot dials people, calling-hour rules and redaction belong in configuration, and your recordings become training data, so check consent and retention before you upload them.

3. One Token to Opt Out of AI Training, Rolled Out at Different Speeds

On October 30, OpenAI, Google, Microsoft and Anthropic said they would honor a single robots.txt token, reported as __usage_ai-training, for opting out of AI training. Timing differs: Anthropic reportedly supports it already, OpenAI says it aims to in the coming months, Microsoft says February 1, 2027, and Google sometime in 2027. Scope differs too. Coverage says Google treats the token as covering both training and Gemini grounding, while Microsoft plans separate controls for each. Enforcement is voluntary and unverified, and opting out may reduce how often content surfaces in AI experiences. Until the signal is honored everywhere, publishers are advised to keep crawler-specific blocks.

If your chatbot fetches or summarizes web pages, read robots.txt and any AI-use signals in your own crawler, and log what you skipped. If you publish content, decide whether you want out of training, grounding, or both before choosing a setting.

Conclusion

The thread is who is on the other end and what they agreed to. Customers' agents need identity and limits, outbound voice bots need rules baked in, and web content now carries a machine-readable preference that vendors honor on different schedules. Build for all three as explicit policy, not afterthoughts.

Suggested visuals: a flow diagram routing human versus agent callers to different tool scopes, and a rollout timeline for the opt-out token by vendor.

— Maya