about agent-teams

[boxiaolanya2008]

Question

When will the agent-teams feature be added in Opencode?

[github-actions]

This issue might be a duplicate of existing issues. Please check:

• [FEATURE]: Add Agent Teams Equivalent or Better #12661: [FEATURE]: Add Agent Teams Equivalent or Better
• [DESIGN]: Agent Teams — flat teams with named messaging, multi-model support, and TUI integration #12711: [DESIGN]: Agent Teams — flat teams with named messaging, multi-model support, and TUI integration
• [FEATURE]:TEAM OF AGENTS #12744: [FEATURE]:TEAM OF AGENTS

These issues already discuss the agent-teams feature. Consider adding your thoughts or upvoting there instead of opening a new issue.

[OS-Lihua]

Which is the core issue regarding adding Agent Team mode?

[boxiaolanya2008]

Which is the core issue regarding adding Agent Team mode?

Honestly, the core issue is speed. One model doing everything sequentially feels sluggish. I'd love to see multiple agents run in parallel—like one writing code while another plans the next steps—to make the whole workflow faster.

[psaawkel]

It's not only faster - it would be nice to make a strategy - i mean to assign roles with specific models - eg. opus will do a research, grok could be a manager, sonnet/GPT - implementation (in parallel), gemini - verification. It could make process cheaper, more cost effective, predictable.

[jensenojs]

• [Feature Request] Swarm tools like "memory system" tctinh/agent-hive#17 (comment)

However, this scenario can actually be handled by subagents (or parallel subagents). One unique value of an agent team, in my opinion, is the ability to focus on problems that do not yet have clear answers. Let agents engage in red team-blue team exercises. As the facilitator, we can direct their attention to a specific domain.

[lrq3000]

I think the Agent Teams as implemented by Anthropic is a big exciting step forward in multi-agents orchestration (and I don't even have a subscription with them and I am certainly no fan of their work).

Claude Code Agent Teams are not merely "parallel subagents." Anthropic defines them as multiple Claude Code sessions where one acts as a lead, teammates (ie, the subagents) work in their own context windows, teammates can message each other directly, and coordination happens through a shared task list. So each has their own context window + a shared one (the shared task list). It's also possible for the human operator to interact and course correct with any teammate at anytime. Teammates can work in parallel or wait for others to finish or correct each others outputs.

This design choice is far superior to subagents, as it introduces far more degrees of liberty and feedback loops in the system, which evidently makes it more costly to run, but it allows to run complex workflows OR automatically decompose vague prompts with huge implications into an autonomous actionable workflow for AI models that were not even possible before.

Also, given how the architecture works, I expect it to work especially well with smaller reasoning models that are capable of advanced multisteps tools calling such as the Qwen3.5 series.

I am convinced that whoever will implement this will see two main benefits:

1. a huge successful completion increase for complex coding problem that were not even possible to tackle before.
2. my personal hypothesis: this boost should be especially noticeable with SLM models such as Qwen3.5, and hence for local use. This is a design that IMHO scales to any model size.

[archichen]

we need team mode, team mode could produce high quality jobs, and like Claude code, user could create a team in instruction would be very easier