Main changes in v0.2.26

• Pi agents on an Amazon Bedrock connection work again: they list models, open old sessions and run schedules.
• Saving a model in the Model settings panel works again on providers such as the IBM proxy. The panel now offers only the models the provider has.
• Codex agents on Azure DeepSeek no longer fail on their first message.
• Agents with two GitHub accounts can download from public GitHub again.
• Skill installs no longer fail with "skill-ref failed" during a brief connection outage.
• Spotlights kit runs complete again.
• A piped dam chat no longer crashes the agent.
• Agents running in kit images with their own harness get their model list again.
• In Slack, an agent that reacts to a message no longer also replies to it.
• The kit update bar goes away once an update finishes, and shows progress while it runs.

Other improvements

• Restricted artifacts can have up to 100 viewers, up from 50.
• A Claude Code call on Bedrock is now counted once in Spend, not twice.

Fixes after v0.2.24

• Schedules and Slack replies work again for agents whose default model ended up "not set" after the upgrade. Bob agents now default to Bob's fast model.
• Sessions from before v0.2.24 show their titles again instead of "(no title)".
• A single bad message from an agent no longer drops everyone's open chats.

Virtual machines for agents (rolling out gradually)

• Agents can now run in their own lightweight virtual machine, which can run Docker and k3s and wakes up faster. We're moving agents over in stages; nothing changes for you until your agents move.
• Once your agents run on virtual machines, they share one memory pool, so you no longer size each agent. Home shows how much of the pool you're using.
• Idle agents on virtual machines go to sleep after 5 minutes. You can change this in the agent's settings.

Other improvements

• Spend shows what was actually charged. Calls your subscription covers no longer count as dollars, and shell runs and Bedrock calls are now counted.
• dam update updates the CLI.
• dam chat and the web terminal reconnect to the same session after the connection drops.
• GitHub App connections are checked when you save them, and Custom MCP servers with OAuth work with more providers.
• A schedule that fired at a sleeping agent no longer loses its run, and schedules with set hours fire at the right time.
• In Slack, agents can read files from earlier in a thread, and the Stop button stops the agent. Slack and Telegram sessions are now named after the first message, and artifacts are shared as links.

Claude Code agents keep going in auto mode

• After a tool call is denied, Claude Code agents now try another way instead of stopping for good, as they did before v0.2.20. You no longer need to add allow rules by hand to work around this.
• On the ETE LiteLLM proxy, safe tool calls are no longer wrongly denied in auto mode. Existing connections get the fix automatically.
• Claude Code is updated to 2.1.293.

New model provider: Curve Bender

• You can connect Curve Bender as a model provider, alongside the ETE LiteLLM proxy. It serves open models (GLM-5, Nemotron-3 Ultra).
• It works with Claude Code, Codex, Pi and Bob agents, and with the 7 starter kits that already accept the ETE proxy. An agent can have both connections.

Pick any of your provider's models in Claude Code

• On a LiteLLM provider (Curve Bender, IBM LiteLLM), the agent's Config panel now lists all of the provider's models below the standard tiers, the same list you see in /model.
• The model list now loads on agents where it used to stay empty.
• If you switch an agent back to Anthropic or Bedrock, the leftover LiteLLM models disappear from the list.

Codex and Bob agents chat again

• Codex agents no longer fail to start a chat with "agent process is not running".
• Bob agents no longer crash on startup, in chat or in the terminal.

Other improvements

• Claude Code agents keep their usual context window. They no longer switch to the 1M window on their own.
• Pi chats on Curve Bender no longer mix reasoning text into the answer.
• Slack whispers on installs with several connected workspaces no longer end after the first reply. A whisper opened before this release may still end once. Opening a new one fixes it.

Schedules work again

• You can create and edit schedules again. Saving a schedule no longer fails with "Failed to load timezone info".
• Existing schedules run on time again.

Delegate to sub-agents (#3419)
Agents can hand off tasks with a tool call, choose the model each sub-agent runs on, and show what each one returned and what it cost, right in the chat.

Starter kits (#3576)
Kits are now the default way to create an agent. Agents also pick up later improvements to their kit.

Slack (#3640, #4166)
You can connect several workspaces. Anyone can talk to an agent privately with /dam whisper (#4300). Agents only post in channels their owner is in, and conversations are more reliable.

New connections
AWS Bedrock as a model provider (#4207), S3-compatible storage where the agent never sees the keys (#4024), and two accounts of the same service on one agent (#3323).

Artifacts (#3096, #2884)
You can edit an artifact's text in place, and interactive artifacts can load and save data through their agent.

Schedules (#3851)
Run now fires a schedule right away.

Changes to know about

• Experiments has been removed (use sub-agents instead).
• dam agent create now requires --provider (#3786).
• To have an agent post in Slack channels, its owner needs a linked Slack account (/dam login).

Share and organize your artifacts

• Share artifacts with specific people by email. Only invited people who sign in can view them.
• Browse artifacts by folder in chat and on the agent page. Drag artifacts into folders directly from chat.

More choice for knowledge bases

Create knowledge base agents with Claude Code, Codex, Pi, or Bob, and choose from the available model providers.

Manage your agent’s resources

Choose a size of 1x, 2x, or 4x in agent settings. A new compute meter shows your available capacity, with clearer guidance on freeing it up.

Run agents from scripts

Use dam run to give an agent work from the command line. Watch results live, run tasks in the background, or retrieve results later.

Other improvements

• GitHub setup now brings sign-in, personal access tokens, and your own GitHub App into one place, including GitHub Enterprise.
• Queued messages survive page reloads. Retry, copy, or delete messages that weren’t delivered.
• Agents can attach files in Slack thread replies.
• With Session costs enabled, Usage shows spending by activity type.
• Skills now show which platform feature they belong to.
• Agents can automatically resume work interrupted by memory failures or unexpected restarts.

Interface change: Experiments is hidden from navigation. Existing experiment agents remain accessible under Coding agents.

New agents work immediately. A newly created agent could sit for 10–25 minutes with no model access — every prompt failing — while being perfectly healthy: its configuration was queued behind a single unrelated agent that had stopped answering. Delivery is now isolated per agent. A hung delivery is cut off after 60 seconds instead of five minutes, an agent can only ever have one delivery in flight, and capacity is 16× larger — it now takes 128 stuck agents to delay anyone, where one used to be enough.
(#3530)

Agents start faster, and the platform keeps up with itself. The agent image is down from 2.6 GB to ~307 MB. The session list and Files panel update as things happen instead of polling, so a session started by a schedule or a file a running task just wrote appears immediately.
(Epic: Measure and cut platform latency — #3242)

Slack conversations behave like conversations. An agent sees what was said in a thread while it was away instead of going blind after its first reply, and a thought split across several messages draws one answer rather than a partial reply per fragment.
(Epic: Improve the Slack Channels experience — #3091)

Agents reply in Slack again on multi-replica installs — replies were being silently lost. Default agent memory is now 2 GB.

A guided start instead of an empty screen. New users get a welcome and pick how they want to enter DAM. A new Home page shows what's running, what's blocked, what you spent, and what you missed. Each feature has its own setup page, so creating an experiment no longer asks questions meant for a coding agent. Clicking an agent opens it in chat, with model, size, connections and skills right next to the message box. If a model provider is missing, DAM says so up front.
(Epic: First time user experience — #3097)

Artifacts you can organize. Open an artifact full screen from chat, rename it, move it into a folder. Copying a share link now confirms it worked. And the retention setting no longer looks like share-link expiry — it deletes the artifact, and now says so.
(Epic: Artifacts usability — #3096)

Several agents in one Slack conversation. Multiple agents can share one Slack conversation and see what each other said, so a team of agents fits in one channel. Agents recognize when they're tagged, messages arrive in order, and documents you attach reach the agent.
(Epic: Improve the Slack Channels experience — #3091)

The platform is much faster. Sessions open fast. The interface gets changes as they happen instead of polling, so agents, sessions and skills stay current — including changes from your teammates. Agent startup is about 18 seconds shorter.
(Epic: Measure and cut platform latency — #3242)

Skills you set up once. Search your skills, toggle them in bulk, and carry a selection over to the next agent. A failed install now fails visibly and retries work, instead of reporting success and leaving nothing behind.
(Epic: Close the Skills usability gaps — #3022)

Chat keeps your place. Your draft stays with the session it was written for. A session shows when background work is still running, so it no longer looks dead between turns.
(Epic: Chat experience on par with native Claude Code — #3249)

Clearer words. DAM says "agent" where it used to say "sandbox". New agents, experiments and knowledge bases get real names like coding-3 instead of random strings.
(Epic: Improve product copy — #3102)

Bob Shell agents move to 2.0. Bob agents now run Bob Shell 2.0, keep their chat history, and resume conversations natively. Chats created before this release don't appear in the sidebar — 2.0 stores history in a new place. New chats are fine.

Also in this release.

• Every new experiment ships with one example you can run right away.
• Sessions have their own link again, so sharing and browser back/forward work.
• Spend by agent shows real agent names, not internal identifiers.
• Attachment limits say 50 MB, the actual limit.
• Gateway restart warnings appear only when the gateway restarts.
• Opening an agent you don't own shows a public agent page.
• Vulnerable dependencies updated.

What's new in v0.2.13

Skills you can trust. Read a skill before you install it, manage the ones you authored, and see when a source was last checked. The Skills surface now tells you the truth about what's there.

Agents in Slack, for everyone. Channels is no longer hidden behind a flag. Put an agent in a channel and the whole team talks to it — one agent, not a copy per person, so it's faster. An agent can sit in several channels, and you can jump from any Slack thread straight to that conversation in the product.

See what your agents cost. Usage per sandbox and cost per session, where you're working rather than buried in a monthly total.

Agents that start when you ask. Pick a size and it runs — no more settings too small to boot. If your budget is full, DAM parks your own idle agents and starts the one you wanted instead of turning you away.

Tighter access for GitHub. Give an agent only the repositories and permissions it needs, instead of everything your organisation installed.

More frameworks out of the box. SkyDiscover AdaEvolve & EvoX join the catalog, and the evolutionary optimizers now all run the same way.

A calmer product. One consistent look across every page, clearer wording where it matters, and a long list of fixes — interrupted replies stay on screen, back and forward work everywhere, artifacts and connections behave the way you'd expect.