AI agents cross from talking to acting, while identity and trust become the sellable surface
The week in communication-messaging
AI in this sector stopped being a feature you switch on and split into two separate things: an agent that acts, and a pipe into someone else's model. Chatwoot's Captain now mines resolved conversations for the questions customers repeat, hands back ranked FAQs to review, and reports what it deflected — the work has moved from adding capability to proving and governing it. Salesmsg pushed its calling and texting agents past conversation into the contact record, writing post-call notes and applying intent tags automatically. And the mailbox and the meeting are increasingly consumed from outside the product: Superhuman connected every Gmail and Outlook account to Claude, ChatGPT and Codex through its Mail MCP, and Grain became a ChatGPT plugin with an MCP server of its own. The interesting AI surface is no longer the chat box inside the app — it is the agent that writes back, and the protocol that lets an external model read the whole inbox.
The other half of the week is identity, trust, and who pays for it. Bandwidth launched Identity Presentation, rendering brand name, logo and call reason on the handset with a pre-call authentication API so the display cannot be spoofed — it is selling whether a call is worth answering rather than another minute of connectivity. Twilio flipped fraud-monitoring defaults on live accounts in Italy and Argentina in response to SMS pumping, and kept absorbing infrastructure it used to leave to customers. Stalwart added SCIM v2 provisioning to its Enterprise tier; Rocket.Chat moved LDAP and SAML behind a Premium licence for 9.0 while forcing end-to-end encryption on new private rooms. Provisioning, fraud posture and SSO are all being re-owned and repriced at once.
Leaders
Twilio is retiring Functions Classic and migrating every account itself between September and October, with no action required from customers — the clearest instance of the pattern running through its window, where Twilio operates the substrate and moves customers rather than emailing a warning. Automatic SSL provisioning for SendGrid link branding and the forced fraud-posture change in two countries show the same instinct to change live defaults itself. At a velocity score of 8.8 it is the busiest product in the sector this week.
Chatwoot shipped an improved Captain that mines resolved conversations for repeat questions and proposes FAQs ranked by frequency, while the assistant overview finally reports deflection as a measured number rather than an asserted one. The governance framing — per-assistant audience and schedule rules, an effectiveness overview — is the shape of an AI feature being readied for buyers who have to defend deploying it, not for teams trying it out.
Salesmsg led with Reviews, a branded review-collection product folded into the SMS channel it already owns, then extended its AI agents from conversation into the contact record with automatic post-call notes and intent-based tagging. Each release adds one more thing the agent can do without a person; the constraint now reads as customer trust that the agent handled the call correctly rather than raw capability.
Bandwidth launched Identity Presentation in beta, putting brand name, logo and call reason on the recipient's phone with a pre-call Identity Authentication API to stop the branded display being spoofed. It is the first entry in months that adds a product line rather than removing a support ticket, and it points Bandwidth at trust as the thing worth charging for.
Stalwart's 0.16.20 broke a long run of pure spec-completeness work by adding SCIM v2 to the Enterprise tier, so a self-hosted mail server can be provisioned and deprovisioned by an identity provider directly rather than through its own directory sync. The same release closed a Redis Sentinel path where rediss:// URLs reached the master in cleartext.
Wildcards
Misskey deleted its bundled machine-learning stack — nsfwjs, TensorFlow.js and the NSFW model are gone — and made sensitive-media detection an HTTP call to a separate service operators must stand up themselves. Skip that step and detection silently stops, treating every upload as non-sensitive. In a sector adding ML everywhere, shedding it is the off-pattern move.
Superhuman's Mail MCP now connects every Gmail and Outlook account a user has and searches them in parallel from Claude, ChatGPT or Codex, while Auto Drafts 2.0 attaches a ready reply to every message that needs one. Superhuman keeps making the mailbox usable from everywhere except Superhuman.
Apprise declared v1.13.0 the last release of its v1.x line and asked every downstream project to pin apprise below v2 before a breaking major arrives. A notification library setting a deadline for its own ecosystem is off the sector's feature-race pattern entirely.
Themes that compounded
- AI agents crossed from conversation into action — writing CRM notes, tagging by intent, and proposing FAQs mined from resolved tickets (Salesmsg, Chatwoot).
- MCP became the way comms tools get read by external models, with whole mailboxes and meetings piped into Claude and ChatGPT (Superhuman, Grain).
- Identity and anti-spoofing turned into a sellable surface — branded calling and fraud-posture changes on live accounts (Bandwidth, Twilio).
- Enterprise identity got repriced: SSO moved behind a licence and SCIM provisioning added to a paid tier (Rocket.Chat, Stalwart).
- Self-hosted servers kept trading convenience for operability, shedding bundled dependencies or demanding manual migration (Misskey, Dovecot).
Watch this week
Twilio's staged infrastructure migrations land through September and October — Functions Classic moves account by account, and Event Streams gets new regional IP ranges as the legacy shared range heads for retirement — so the retirement-notice cadence is the thing to track, not a new feature. Rocket.Chat's 9.0 SSO licence requirement is the story of its next release rather than any calling capability. And with agents now writing to records at Salesmsg and Chatwoot, the open question is whether acting on a customer's behalf gets the same audit and governance scaffolding that deflection metrics just received — the binding constraint this week is trust, not capability.