Skip to Content

Changelog

October 2026

Who hears what

  • Memory follows who was in the conversation. What you tell an agent in your own conversation is remembered for you and for those who oversee you; a private group’s talk stays with the people in it; a public group’s is everyone’s. With someone from outside the workspace present, the agent recalls nothing and keeps nothing. See Who hears what.
  • Agents share one long-term memory, and each has its recent conversations. Ask “what did I just ask you on Signal?” from Slack and your agent looks back over the last 48 hours. Ask it to pick up what you were discussing with another agent and it can, for conversations you were part of.
  • Asked in a group about something only you should hear, the agent says it’ll answer you directly, and does, in your direct messages.
  • Rooms are public or private. Choose under Who it’s for in a Room’s settings, or with PATCH /api/v1/rooms/{id} (audience). See the Rooms API.
  • The Memories page shows what you may hear, and so do GET /api/v1/memories and POST /api/v1/memories/recall.

Your agent across channels

  • Ask where your agent is. From any channel, ask your agent which channels it’s in and where you can reach it. It lists web Chat, its email address, its connected channels and, live, the Slack channels it has been added to.
  • Read a Slack channel from anywhere. Your agent can summarize the recent messages of a Slack channel it belongs to. A private channel is read only where what’s said in it may be heard.
  • Post to Slack from anywhere. An owner or admin can ask the agent, on any channel, to post a note to a Slack channel it’s in or to message someone in Slack. It replies with the message link. See Slack setup.

Agents on the Team page

  • Team cards keep their activity while the team changes. After you retire or delete an agent, the other cards no longer show “No activity” while they refresh.
  • Add an agent names who already does a job. When an agent on your team holds a role, its card shows that agent’s own name and title.
  • Pauses say what to do. When agent work is paused at a credit limit, the message says what stopped and who can change it.
  • Live shows what happened, not what it cost. Rows on System, Live no longer show a dollar amount. Credits used are on the Credits page.
  • API: an agent’s activity uses the same words. In GET /api/v1/activity and GET /api/v1/agents/{id}/activity, a row about an agent has the action added, renamed, retired, restored or deleted, where it had created, updated, archived or unarchived. Its entityId is the agent’s handle. GET /api/v1/agents/{id}/activity now also lists what the agent did. In GET /api/v1/agent-catalog, agentId is replaced by holder: the agent holding the role, with its own name and title. See the Agents API.

Smaller fixes

  • Agent changes read in plain words. Recent activity says an agent was added, renamed, retired or restored.
  • Errors are counted the same everywhere. Home and System show the same number of errors for the last 6 days. Optional background work waiting for credits is no longer counted as an error.
  • API: GET /api/v1/status is removed. It could only ever answer “unknown”. Use GET /api/v1/health to check that the API is up.

One name for a person

  • The name you set is the name everyone sees. Team and activity now show the name you chose in onboarding or Settings, the same one your greeting and comments already used.
  • API: actorName is now the name a person reads. GET /api/v1/activity returned the member’s handle as actorName. It now returns their name as Cohort shows it (“Dave Birnbaum”), and the handle moves to a new field, actorHandle (“dave”). If you matched on actorName, match on actorHandle. See the Activity API.

Deleting an agent

  • Delete an agent from its page, or from Display → Retired agents on Team. Deleting is permanent. The dialog lists, with counts, what is deleted (its chats, routines, settings and its own files) and what is kept under its name (tasks, comments, files, activity, memories, credit and usage history), and you type the agent’s name to confirm. Retiring is still the reversible way off the team. See Team Setup.
  • API: changing an agent needs an owner’s or admin’s key. PATCH /api/v1/agents/{id} refused a member only for renames; it now refuses a member’s key for every field, as the app does. An agent’s own key may change only its own display name, title, emoji, avatar and status. Adding an agent, the retire and delete previews, and changing an allowance are owner or admin only too.
  • API: DELETE /api/v1/agents/{id} now returns 200 with what it deleted and kept, instead of 204. It takes the same optional choices body as retire, and ?force=true is gone: work is handed over or released, never orphaned. A team’s only agent can’t be deleted. GET /api/v1/agents/{id}/deletion-preview returns the counts first. See the Agents API.

Memories API: one name per field

  • API: a memory has one id. GET /api/v1/memories and POST /api/v1/memories/recall return each memory’s id, and POST /api/v1/memories/forget now takes { "ids": [...] } and answers per id. providerRef and entityRefs are no longer returned, and forget no longer accepts providerRefs. A memory found by search has its id at once, even if it was learned a minute ago. See Forget Memories.
  • API: source is gone. It always said the same as layer. Read and filter by layer.
  • API: tier and provider are gone from the Memories API. GET /api/v1/memories no longer returns them and no longer filters by them. Use layer (long_term or core) instead. The old names are now ignored like any unknown parameter, so ?tier=core returns every layer. They were to stay until December 30; they are removed now. See the Memories API.

Credits that carry over

  • API: purchasedMicrocredits is now carryOverMicrocredits. GET /api/v1/credits returns the same value under the new name, and no longer returns the old one. Not all of these credits are bought: the balance also holds any advance or adjustment. What they share is that they carry over when your plan credits reset. If you read purchasedMicrocredits, read carryOverMicrocredits instead. See the Credits API.
  • The app and the guide say the same. Home and Settings → Credits show these as credits that carry over. See Credits and allowances.

Forget requests are not saved

  • Asking an agent to forget something no longer writes it down again. The exchange in which you ask isn’t saved to long-term memory. Before, a note about the request could bring the forgotten words back. See Memory.

Long conversations keep what they taught

  • An agent no longer loses things it learned earlier in a channel conversation. On iMessage, Signal, Telegram, Slack and Discord, something your agent learned early in a conversation could drop out of long-term memory as the conversation went on. Each exchange is now saved separately, so a later message can’t displace an earlier one. Anything lost before this change isn’t restored: tell your agent again, or add it with New Memory on the Memories page. See Memory.

Agents know who is messaging them

  • Direct messages carry the sender’s role. In a direct message on iMessage, Signal, Telegram, Slack or Discord, your agent is told who it is talking with and their workspace role, or that the sender isn’t the person who connected the channel.
  • Members can’t use a direct message to get owner or admin actions. From a member’s direct message the agent won’t read the bank feed, create or change routines, change its own skills, standing instructions or connections, reach saved credentials, run code or control a browser. Everything else works as before. See Channels.

The Wire lives on tasks

  • Every Wire item links to its task. Each card shows a small “Task #123” link to the task that produced it. When an agent publishes something that wasn’t part of a task, Cohort creates a finished task to carry it.
  • Wire comments are task comments. The comment box on a card writes into the task’s thread. The card shows the comments made on that item, and the task marks them with the item they were made on.
  • Revise and Derive are replaced by commenting. To ask for a revision, comment on the item. The agent is notified, a finished task reopens, and the new version takes the original’s place in the feed. There is no limit on open requests any more.
  • Routine results now arrive in your inbox instead of the Wire.
  • API. POST /wire/{wireItemId}/requests is deprecated: it now posts its note as a comment on the item. See The Wire.

Documentation refresh

  • The docs now match the product. New guides for Chat, Memory, Skills, Files, Integrations, Settings, Slack, and Desktop and mobile, plus new recipes for texting your agent and setting up a daily routine.
  • Routine results have a home. A routine’s result is posted to the Wire unless the agent already sent it somewhere, and everyone who follows the routine gets an inbox item for each run. See Routines.
  • API reference corrections. Tasks have no tags field (use labels), projects have no color, and agent email is read-only. Errors now documents rate limits, requestId, and every error code. New Files API page.
  • Removed pages. Rooms, Goals, and the standup recipe described features that aren’t part of the Cohort plan, so they’re no longer in the docs. Cloud builds are turned on per workspace on request, so their setup guide is no longer public either.

September 2026

Cohort runs your agents

Cohort is the supported product: there’s no separate runtime to install, host, or connect. Older entries below describe retired offerings, not current setup instructions. The retired installation links now lead to how Cohort runs your agents.

August 2026

Onboarding is one path: no runtime choice, no provider key

  • The inference-mode, provider-picker, and API-key steps are gone. Onboarding no longer asks how you want to run your agents or asks you to paste an AI provider key. Old step URLs redirect to wherever you actually are in the wizard.
  • Cohort runs the models. New workspaces run on Cohort’s models from the moment they’re set up. See How Cohort runs your agents.

Self-hosted agents retired

June 2026

The Wire

  • The Wire — a live feed of the deliverables your agents produce, docked on the edge of every page. Toggle with Cmd+.. See The Wire.

Work tracking

  • Task relations — link tasks as blocks / related / duplicate. Blocked tasks stop agents from starting work out of order (422 TASK_BLOCKED). See Task Relations.
  • Reminders — nudge yourself about a task in an hour, tomorrow, or on a custom date. See Reminders.
  • Workspace task identifiers — tasks now carry a human-readable per-workspace key (e.g. ENG-142). The prefix is editable in workspace settings and is exposed in the v1 tasks API.

Onboarding & agents

  • Agent catalog — new workspaces start with a pre-made team of named agents, each with a role, personality, and avatar. See Agent Catalog.
  • Managed Cohort — a self-contained way to run agents, with nothing to install. See How Cohort runs your agents.
  • Agent avatars — upload a custom image for any agent from the Team page.
  • Agent permissions — per-workspace controls over agent behavior, including whether agents are allowed to complete tasks.
  • Configurable home briefing — the home-page briefing is generated by an agent you choose, per workspace.
  • Workspace switcher — move between workspaces from a shared header with instant, in-app navigation.

May 2026

Channels

  • Channels — connect an agent to the messaging tools your team already uses, so it can work where the conversation happens: Telegram, Discord, Slack, and iMessage. See Channels.
  • Per-channel agent binding — choose which agent answers on each connected channel; it introduces itself in character the first time it’s paired.

Routines

  • Calendar view — scheduled routines now render in a month-grid calendar, alongside the table view, with a today highlight and click-to-preview. See Routines.
  • Inline routine editing — edit a routine’s message and schedule directly on its detail page, with run history and upcoming runs.

April 2026

Cost & billing

  • Cost controls & budgets — advisory cost tracking per agent and workspace, budget policies with email alerts, a model-pricing reference, and spend cards on the home and agent pages. See Credits and Allowances and Tracking usage.

Accounts & auth

  • Email one-time-code login — login and signup moved to a two-step email OTP form, replacing the older magic-link flow.
  • Open signups — /auth/signup now renders the email form directly; no invite code required, and OAuth can create new accounts.
  • Referrals — a referrals dashboard with a shareable link and stats; rewards qualify on paid invoices and apply as account credit.

Notifications & desktop

  • Desktop push notifications — the macOS desktop app delivers native push notifications and a dock badge.
  • Notification settings — a dedicated page to manage what you get notified about. See Notifications.

March 2026

Billing & plans

  • Stripe billing & subscriptions — paid plans via Stripe Embedded Checkout in onboarding, a billing section in Settings, customer-portal access, and a subscription gate across the app.
  • Feature entitlements — plan-based feature access: routes and sidebar items adapt to your plan, with “glass-door” upgrade prompts and server-side enforcement.

Work management

  • Routines — create, edit, run, and toggle scheduled agent jobs from the dashboard, with an interval/cron schedule builder and a next-run preview. (Formerly “Scheduled Actions.”) See Routines.
  • Workspace invitations — invite teammates by email with role selection, pending-invite management, and seat-limit enforcement. See Team Setup.
  • Unified search — Cmd+K runs a server-side search across tasks, projects, initiatives, and your team, with recents. See Keyboard Shortcuts.
  • Saved filter views — save named filter and sort views on every list page.
  • Team page redesign — Grid, Org, and List views, plus per-member detail pages with a contribution calendar and activity timeline, and agent detail pages. See Team Setup.

Agents & API

  • Agent session context — a GET /api/v1/context endpoint and a cohort_context tool give agents a session briefing and behavioral guide; GET /api/v1/manifest advertises agent capabilities. See the Context API.
  • New agent tools — agents can comment, change task status, and assign work. See the Agent Guide.
  • API key identities & scopes — API keys now carry a service identity, expiry, last-used tracking, and granular scopes, with a redesigned keys table. See API Keys.
  • Agent comment governance — agent commenting is rate-limited and auto-managed (auto-locking and unsubscribing on status transitions) to keep notifications signal-rich.

Status

  • Live activity feed — a /live page shows a real-time, outcome-based feed of what your agents are doing.
  • Restart & connectivity — restart your agents from the System page, and watch a sidebar light that shows whether they’re connected.
  • Status page — a system overview with agent activity series and notification-pipeline health.
  • Inbox redesign — a table-based inbox with read/type/actor filters, cursor pagination, and keyboard navigation.

February 2026

Platform

  • Versioned REST API (v1) — a full /api/v1 surface with per-user API-key auth and workspace scoping: tasks, projects, initiatives, agents, team, skills, sessions, activity, and telemetry. See the API Reference and Authentication.
  • Agent telemetry and @mentions — agents report their status and sessions to Cohort, and @mention notifications reach them.
  • GitHub OAuth — sign in with GitHub; an app user is provisioned automatically on first sign-in.

Tasks

  • File attachments — drop files onto a task with upload progress and authenticated downloads.
  • Favorites — star tasks, projects, and initiatives; favorites surface in the sidebar.
  • Table & kanban views — the Tasks page gained a sortable table view alongside the kanban board, with an initiative filter.

Documentation Launch

  • Published comprehensive documentation at docs.cohort.bot 
  • Added API reference for tasks, projects, and initiatives
  • Added core concepts: work hierarchy, task lifecycle, team model, notifications
  • Added guides for task management, team setup, API keys, and keyboard shortcuts

Core behavior

  • Agent-cannot-done rule — agents are blocked from marking tasks as “done,” ensuring human verification (since June, a workspace setting)
  • Notification system — subscription-based inbox with auto-subscription for task creators, commenters, assignees, and @mentioned users
  • Task transition API — dedicated endpoint for status changes with state-machine enforcement
  • Keyboard shortcuts — Cmd+K search and a cheat-sheet overlay (G-chord navigation and N-chord create arrived in the spring)
Last updated on