AgentTeams 1.2.8: Unified Workbench Supports Six Harnesses for Multi-Agent Collaboration
AgentTeams 1.2.8 introduces a redesigned workbench unifying six agent harnesses, enhanced tracing for model and tool calls, voice input with editing, and improved team collaboration workflows that increase delivery success rates while reducing context overhead, alongside a new website and one-line installation.
AgentTeams is an open-source multi-agent collaboration system where a Manager coordinates a team, Leaders break down and verify tasks, and Workers execute work. Different agents can use different runtimes and models. Collaboration messages flow through Matrix rooms, files are stored in shared storage, and model/MCP tool access is managed by the Higress AI gateway. The Dashboard presents these capabilities in a single workspace for tasks such as code development, research, and content production.
Version 1.2.6: Natural Input, Visible Execution
This release focused on answering: after a task is delegated, how do you know what the agent is doing?
Voice Input, Edit Before Sending
Chat supports Chinese/English speech-to-text using browser-local recognition by default; the model is cached after first load. Recognized text enters the input box first, allowing wording changes or added constraints before manual send. Admins may also configure Bailian speech recognition.
Execution Tracing: See Model and Tool Latency
Optional Worker native session tracing was added, correlating QwenPaw and Hermes execution rounds, model requests, and tool calls. The Dashboard provides a session list and execution waterfall view to identify where time is spent and which step failed. For example, a stalled research task can be diagnosed by checking whether the agent is waiting for the model or a tool request failed.
Human Intervention: Explicit Items and Records
TeamHarness gained pause/resume control for task dispatch and loop progression with operation logging. A team-level human-task inbox centralizes confirmation requests, and human replies are stored structurally. Model authorization, team permissions, personal file areas, and storage sync were also refined, clarifying who can configure models, manage resources, and access files.
This version suits teams already running collaborative tasks who need visibility into execution and human intervention points. Native tracing requires separate backend configuration and enablement; runtime coverage depends on harness support.
Version 1.2.7: One-Click Start, Six Harnesses in One Team
1.2.7 further lowers the entry barrier: start the workspace first, then configure the Manager model in the Dashboard. Data-preserving upgrade flows were also improved.
Six Harnesses, Unified Image and Versioning
OpenClaw, QwenPaw, Hermes, DeepSeek Harness, OpenHuman, and Codex now share a common harness image naming and versioning scheme, providing both amd64 and arm64 images. Manager and Worker reuse the corresponding harness images, reducing the number of distinct image types needed at deployment. Users can select different runtimes per task and organize agents with diverse strengths into a single collaboration system.
Tracing Extended to General Agent Conversations
Diagnostics are no longer limited to formal task scenarios. General tracing supports queries by agent, team, session, task, and time range. The Dashboard workbench adds a tracing entry; the Controller centralizes configuration. Even a single direct conversation with an agent can be inspected for model requests, tool calls, and failure steps. This still requires a configured collector and observability backend.
Multi-Team Daily Use Improvements
Team, Manager, and unassigned Worker management entries are more centralized; team topology and agent details are viewable in-page. Model selection, human to-dos, personal file sharing, and first-login guidance continue to improve. Chat groups thinking and tool calls per round, shows real-time previews during execution, and auto-collapses afterward. Project completion conditions are stricter: a project can only be marked complete after all tasks are finished and deliverables recorded.
Version 1.2.8: New Dashboard and Team Collaboration Upgrade
The most visible change is the overall workspace visual and interaction refresh: white background, dark gray text, thin borders, and yellow-green primary actions concentrate attention. Home, Assistant, Task Board, Team, and Assets pages adopt a more consistent layout and operation style. More importantly, the flow from requirement handoff to viewing assignments, entering collaboration, handling approvals, and checking results is now smoother.
Workspace Home: See Progress First, Then Enter Tasks
The new home page, themed "One-Stop AI Workbench: Build-Collaborate-Govern," aggregates in-progress tasks, collaboration teams, and workspace assets. After entering the workspace, users can chat directly with the AT Assistant or open a task to continue tracking. Global search covers pages, tasks, agents, skills, and tools, reducing multi-page navigation.
Chat and Tasks Together for Easier Context
In assistant sessions, a sidebar shows related tasks with access to files and topics. Viewing a task retains the chat context; users enter the corresponding collaboration room only when they need to participate in execution. Task details, dependency graphs, and activity logs are now unified across the task board and collaboration sessions — opening the same task from different entry points shows consistent information.
Team Collaboration Upgrade: Higher Delivery Success, Lower Time and Context Cost
1.2.8 optimizes Manager-Leader-Worker coordination around task dispatch, execution, progress sync, and result acceptance:
More reliable delivery. Main tasks, execution subtasks, and deliverables have clearer associations; dispatch, submission, acceptance, and revision link tightly. Leaders organize acceptance around explicit task states and delivery results; steps needing extra info or human confirmation are easier to discover and handle.
Faster progress. Manager focuses on requirement clarification, team selection, and result reporting; Leader handles breakdown, dispatch, and acceptance; Worker concentrates on execution. Combined with task status and collaboration notifications, this reduces repeated progress chasing, multi-layer relay, and unnecessary coordination waits.
Leaner context. Task requirements, execution progress, and delivery files are organized by main and subtasks. Agents fetch only information relevant to their role, exchanging results via structured progress and shared files, cutting redundant long-history passing and re-reading of irrelevant context.
Main tasks can advance independently or be grouped into projects. Leaders break down and dispatch subtasks; Workers submit results, then enter acceptance or revision. Users see assignments on the board and track actual delivery in collaboration sessions. For example, preparing a product beta test: the user states the goal and acceptance criteria; the team separately compiles interview summaries, drafts invitation copy, and prepares an acceptance checklist. The board shows main task status and subtask completion; confirmation items return to the collaboration session for handling.
The task board places "To-Do, Dispatched, In Progress, Pending Acceptance, Done" in the primary flow, while "Blocked, Needs Revision, Failed" are grouped for quick attention. Task cards link into collaboration sessions to view execution context and delivery status.
Adjust Direction: Stop and Supplement Instructions
During direct agent chat, users can issue stop or supplemental commands for the current execution. For instance, if an agent is compiling an overly broad report, the user can add "focus on deployment and tool integration first" to steer subsequent work. The stop state waits for the native runtime to finish cleanup; already-performed external actions are not auto-rolled back. This interaction targets human-agent direct sessions; team-delegated tasks still use the task control flow.
Approvals are closer to actual execution: native approval requests from all six harnesses link to Dashboard cards, with results updated via runtime callbacks. Member read status ties to mentioned members, helping judge if a message was seen; read status and task completion are displayed separately.
Model Configuration and Tool Access: More Complete Operational Paths
Model management centers on "model connections": enter platform, API endpoint, key, and model ID, then run a short generation request to test connectivity, viewing result and latency. The test checks the upstream model connection; agent authorization and actual tool calls are still verified at use time.
MCP tool access is further completed: connect to existing MCP Servers, or ingest from HTTP APIs or databases. The gateway then discovers tools, allows debugging against parameter schemas, and configures Worker authorization. Database ingestion requires the gateway to enable corresponding data-plane capabilities.
Team Management and Runtime Diagnostics Continue to Improve
The team page supports switching between card and topology views for easier member relationship, configuration, and status inspection. Runtime monitoring cards, container logs, Manager diagnostics, and associated evidence entry points move issue localization from "no reply" to concrete runtime states.
Completed task sessions support archiving and restoration, preserving history messages and artifacts, reducing the burden of ongoing message sync for historical tasks.
Additionally, 1.2.8 introduces controlled Worker harness switching, task checkpoints, and recovery isolation/resumption mechanisms. Some real-world resumption scenarios have been validated; full cross-harness recovery scope is still being refined — no guarantee of lossless switching between arbitrary runtimes.
New Website: From Demo to First Task
The new AgentTeams website is live at https://agentteams.cloud . Users can watch the new Dashboard demo, then install, read docs, or join the GitHub community from the same entry point.
One Command to Launch AgentTeams and Dashboard
Prepare a Docker-compatible container runtime and a usable model API key, run the install command for your OS, then follow the wizard to complete configuration.
macOS / Linux:
curl -fsSL https://agentteams.cloud/install.sh | bashWindows PowerShell:
iwr -UseBasicParsing https://agentteams.cloud/install.ps1 | iexAfter installation, open the wizard-provided Dashboard URL, configure the Manager model, and give the assistant a first goal. For example:
Help me form a research team to compare three open-source knowledge-base solutions. Investigate deployment requirements, retrieval capabilities, and maintenance costs for each, and produce a sourced comparison report. Ask me to confirm any conditions before proceeding.
Then visit the task board to see assignments, add constraints in the collaboration session, and review the final delivered report — a complete path to experience the new Dashboard.
The install script deploys the website-recommended version. For existing deployments, back up configuration and persistent data before upgrading, then follow the upgrade wizard to retain accounts, model configs, and workspaces.
Come Use It, Come Build It
AgentTeams aims to let agents of different runtimes collaborate in the same team, while keeping humans able to see the process, participate in decisions, and verify delivery.
Welcome to start with a real task. If you have new harnesses, tool integrations, collaboration scenarios, or encounter issues, please submit issues, join discussions, and contribute PRs on GitHub. Starring the project to follow future releases is also appreciated.
Website & Demo: https://agentteams.cloud
GitHub Main Repo: https://github.com/agentscope-ai/AgentTeams
Issue Tracker: https://github.com/agentscope-ai/AgentTeams/issues
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.
