Fleet agents delete files; MCP OAuth fixed; Slack files ingest

Fleet

• In the Agent Builder view, the footer workspace and tenant list is sourced from the Fleet API so you can switch between your Fleet workspaces.

• The Access Profiles dialog in chat now includes a Create an access profile link that opens the sandboxes create flow, so you can add a profile when a workspace has none configured instead of hitting a dead end.

• Fleet agents can now delete files from their memory and skills using the new delete tool, including files in linked workspace skills. Core agent files and read-only system skills remain protected.

• Fleet now completes OAuth for MCP servers whose authorization server requires client-secret authentication at the token endpoint, so connecting these servers no longer fails after the consent step.

• First-time Fleet users now see a streamlined welcome modal with two clear paths — describe an agent to build with AI (starting from a prompt in Chat) or start from a curated template — replacing the previous multi-step setup wizard.

• Creating an agent from a Fleet template now skips the setup wizard and opens the agent editor with the template onboarding card.

• Fleet now sends the MCP protocol version a server negotiates during the handshake, both when loading tools and when the agent calls them, so MCP servers that require a newer version no longer return zero tools or fail tool calls.

• Fleet agents receive the day of week alongside the current date (for example “Monday, June 29th 2026”), so scheduling and date reasoning no longer relies on the model inferring the weekday from the ISO date.

• File edits in Fleet agent chat now render as syntax-highlighted, line-by-line diffs, making changes easier to review.

• Fleet agents can now read files shared with them in Slack. Attach an image, PDF, audio, video, or text file in a mention or DM and the agent ingests it into the conversation.

• On the Agent Builder Integrations page, searching now selects the All tab so results span every category, and switching category tabs clears the search.

• When you connect a custom Slack bot to a Fleet agent, Fleet sends the installer a direct message with quick setup tips, including how to add the bot to channels and mention it with @.

• Fleet agents now have a Slack tool for listing channels the connected bot is a member of, making it easier to discover the right channel before posting or reading messages.

• Fleet OAuth provider and integration responses now include an owner field (workspace or platform) so you can tell your own resources apart from built-in, platform-managed ones. The platform manager organization can now create and modify built-in OAuth providers.

• Setting up a schedule is now clearer: choose a preset (daily, weekly, monthly, or every few minutes) or enter a custom cron expression, with a live human-readable preview and inline validation as you go.

• When registering an integration OAuth provider for headless connections, http:// redirect URIs are now accepted only for the loopback IP literals 127.0.0.1 or [::1]. The localhost hostname is no longer accepted over http — use the loopback IP literal or https.

• The MCP servers settings page now scrolls when the pointer is over the servers list.

• The load previous conversations tool now writes conversation files into the attached Computer sandbox when one is enabled, so agents can inspect the downloaded history with their normal file tools.

• When a Fleet agent’s subagent calls a tool that requires human approval, the approval prompt now appears in the chat instead of the run completing without it.

• The Executive Assistant template can now deliver its daily brief and answer @mentions in Slack after you connect a Slack workspace, and both the Executive Assistant and Software Engineer templates received configuration fixes.

• You can now type and send a message in agent chat while a human-in-the-loop prompt is pending. Sending a new message dismisses the pending request and continues the conversation instead of leaving the composer locked.

• Empty sections in the agent configuration panel — Channels, Connections, Skills, Schedules, Instructions, and Subagents — now explain what each one is for and what you can add before you connect anything.

• Creating a new agent no longer fails with a contentBlocks.push error when the chat stream returns string message content.

• Opening an agent in the chat inbox no longer issues repeated duplicate background requests while choosing which thread to open, reducing flicker.

• Fleet agents now load your workspace’s private skills. Previously, in workspaces with fine-grained access controls, an agent could start with only public skills available.

• Reloading an agent chat page no longer flashes the thread list through loading and loaded states multiple times. The sidebar now waits for agent scope to finish loading before fetching threads, so the list settles once.

• GitHub App installations now sync through the authenticated LangSmith session after installation completes, keeping workspace linking aligned with the active user.

• OAuth providers now accept an optional default redirect URI (default_redirect_uri). When set, headless OAuth flows for that provider return the authorization code to it instead of the LangSmith callback, without passing a redirect on every request. The value is validated against the provider’s allowed redirect URIs.

• Fleet agents now discover tools with find_tools or an /tools listing before opening a tool’s reference doc, so they no longer waste a turn reading guessed tool filenames that do not exist.

• The Fleet Fast model tier (gpt-5.4-mini) now runs at medium reasoning effort instead of low, improving response quality on harder tasks.

• The templates gallery now features the Executive Assistant and Software Engineer templates as large cards with a hero illustration, each showing the agent’s own icon.

• Each tool inside a connection in the agent Configure panel now has a remove action (a trash button revealed on hover, matching the connection remove) instead of an on/off switch. The switch implied a reversible toggle, but turning a tool off actually removed it from the agent — so the control now reflects what it does.

• Sending a chat message while clarifying questions were pending could fail the run and leave the thread stuck. Free-text now correctly dismisses the pending request before continuing.

• In the Agent Builder chat, the Skills block’s “Add skill” menu now opens the browse-workspace, create-skill, and import-from-URL dialogs. Previously choosing an option changed the URL but nothing appeared.

• Opening an agent in Fleet now always starts a new chat instead of jumping into a recent thread. Past conversations remain available in the thread sidebar.

• When an agent created from a template introduces itself, it writes what it learns straight to its own memory instead of pausing for approval on every file. Memory writes in your other threads still ask first.

• Skill descriptions containing quotes, colons, or multiple lines are now parsed and stored correctly, and importing or editing a skill preserves all of its frontmatter instead of dropping fields like license or allowed-tools.

• The Add connection dialog now groups Arcade MCP servers under a dedicated Arcade section, so they are easy to find instead of being listed under Other.

• The Fleet model picker now groups served, LCU-billed models (Fast, Pro, Max) separately from bring-your-own models billed per run, making the pricing model for each option clearer.

• The compact Fast/Pro/Max model picker in Agent Builder now shows the model icon on its closed trigger, matching the full model picker.

• When an organization reaches its monthly Fleet usage limit, the error now directs users to upgrade their plan to continue.

New features

• You can now add any agent to Slack in one click. After you authenticate with Slack once, Fleet automatically creates a Slack app configured with the agent’s name, description, and icon, and maps each agent to a single Slack app.

• When an agent is first added to a Slack workspace, it sends the creator a direct message with tips for inviting it to channels and mentioning it.

• Agents now raise tool approvals directly in Slack, with Approve and Deny buttons in the thread, so you no longer need to switch to the Fleet UI to respond.

• When an agent encounters an error during a run, it now replies in the Slack thread instead of going silent. Authentication errors and some other error types include more detail.

• Agents can now read file attachments in Slack messages.

• The agent editor is now a sidebar built into the agent chat page, which organizes configuration into Channels, Connections, Knowledge, Schedule, and Advanced settings drawers.

• The agent creation experience now starts from a blank-slate agent that configures itself and pauses at key points to bring you into the process.