release notes

Smart Node in August 2026: One Person Across Every Channel

August 2026 was Smart Node's first month with numbered releases, starting with 1.0.1 on 2 August and ending with 1.12.22 on 31 August. Over the month, the system shifted from managing separate communication tools to running an office that recognizes people across channels, answers live by voice, and adheres to strict contact boundaries.

One person across every channel

Before version 1.12.0, every channel treated the same person as an isolated record: an account name in web chat, a Telegram chat, or a mailbox address. An agent in web chat could not see what a client had discussed on Telegram without running a long chain of tool lookups.

From 1.12.0 onward, each person has a single record. Agents recall an individual's history across all channels directly from context instead of searching for it. The voice line shares the exact same recall. Recent interactions from every channel appear together so that an active chat session does not drown out a quiet email thread. In addition, all agents now share a single search tool.

Live voice conversations with agent colleagues

The voice line routes to whichever agent is assigned to voice. When a speaker asks a question that requires input from another agent, the assigned agent consults that colleague and returns the answer while the conversation stays open.

Voice conversations take place in the speaker's language. Agents on the voice line remember people, read assigned reference documents, and answer directly from them. Later in the month, voice became the first channel recorded turn by turn. If a node restarts or encounters a crash during a conversation, what was said up to that moment is saved, and ongoing conversations can be searched in real time.

Group email and contact permissions

Until 1.12.18, messages were restricted to a single recipient. Nodes could not reply to a thread, copy a colleague, or loop an owner into a discussion. Messages now accept multiple recipients. Every address is validated using the same checks as before: if any single recipient is refused, the entire message is blocked, and the refusal specifies who caused it rather than silently dropping delivery. Blind carbon copies are intentionally omitted for now.

Rules governing who agents contact were also centralized:

  • An agent only writes to granted contacts and only wakes for granted senders. Whole domains can be approved, but only through an explicit setting.
  • Forwarded email routes to the specific mailbox where it landed.
  • Rules now live in the node registry as numbered lines with the owner's stated reasoning attached. Editing a rule updates every agent that relies on it.
  • Agents are told their own names, their assigned memory boundaries, and the colleagues they can contact.

Built-in workflows and spending limits

Version 1.12.14 added the runtime folders required to run multi-step workflows on nodes built from a release. Nodes include a reference workflow for drafting, an independent review step that can return work once for revisions, and final delivery. Workflow steps run with the node's authenticated credentials, and the visual workflow diagram reflects branches, parallel steps, and loops.

To keep operational expenses bounded, version 1.12.22 lets operators set and adjust spending limits directly without editing the database. An hourly supervisor process checks for approvals that have stalled and sends reminders.

Fixes from live nodes

A substantial portion of the August updates resolved issues reported by operators running active nodes:

  • Agents could not read their own inboxes, making the inbox appear empty. Fixed in 1.12.7.
  • The background watcher that wakes an agent for incoming mail was refused on every poll, leaving messages unanswered. Fixed in 1.12.8.
  • Malformed messages no longer block an agent from reading subsequent mail; the invalid item is skipped and logged.
  • The activity feed on the dashboard was adjusted to show genuine office tasks instead of routine polling checks.
  • Agents account for their own wake events and spending, with cached tokens included in the count.
  • Reading long documents no longer terminates the session, and large prompts no longer stop agents from waking.
  • Notes added to agent knowledge are no longer silently dropped, and write failures report errors.

On release maintenance, version 1.10.5 reduced installation to one command with checksum validation. Releases cannot be published unless a clean node builds from scratch (added in 1.12.6) and passes automated chat tests (added in 1.12.22). One known issue in 1.12.22, which logged model credentials to the system journal during agent runs, was resolved in 1.12.23.

Full notes

Complete technical details for each update are available on the August changelog. Current releases are tracked on the changelog, and the following month's work is covered in Smart Node in September.