You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
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.
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.
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.
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:
a huge successful completion increase for complex coding problem that were not even possible to tackle before.
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.
Teams is a fun new development, but what is an actual problem forcing me to switch to some other framework is the complete inability to spawn multiple subagents at the same time. I sadly cannot wait any longer. It would be a minor change in prompt.ts to allow that.
@allarac I thought OpenCode could spawn multiple parallel subagents? At least OpenCode Server appears to allow that. What would need to change in prompt.ts ? If you make a PR this may get merged soon enough.
@allarac I thought OpenCode could spawn multiple parallel subagents? At least OpenCode Server appears to allow that. What would need to change in prompt.ts ? If you make a PR this may get merged soon enough.
@allarac I thought OpenCode could spawn multiple parallel subagents? At least OpenCode Server appears to allow that. What would need to change in prompt.ts ? If you make a PR this may get merged soon enough.
What are you talking about. It cannot.
I based my assumptions on the fact Kilo Code was rewritten to use OpenCode Server mainly to support parallel tools calls and parallel subagents delegation. So it seems that OpenCode Server is not a blocking point and can support parallel subagents delegation with little changes, if it's not already.
When I use OpenCode it also seems to me it can call and wait for multiple subagents in parallel (using oh-my-openagent), or am I mistaken ?
@panwar-stack Thank you very much for sharing your work! This is very impressive indeed! I saw the HTML5 game prompt, log and result, it is incredible that you got such a complex fully functional game in one go! You confirm that you inputted just the one prompt you uploaded without any followup?
And yes, agents are spawned and working in parallel!
Given the focus of oh-my-openagent, I am not sure agent teams will get merged, or maybe not anytime soon. Maybe you could consider releasing it as a separate plugin for opencode? I for sure will install and try it (but I will anyway clone your repo to try whenever I have some spare time!).
I've finally moved to Ensemble, is a nice approach https://github.com/hueyexe/opencode-ensemble and I've also created a simple CLI called opencode onboard that helps developers (mostyl to myself) to add Opencode + Openspec + Ensemble + some other small configurations to brownfield that have no ai configuration installed and need to follow a structure without adding opinionated skills about development https://github.com/CKGrafico/opencode-onboard
You confirm that you inputted just the one prompt you uploaded without any followup?
@lrq3000 yes, that was all the result of just that one prompt. To be honest, I was not expecting it to work so well on the first go itself. I did have to tweak the implementation so the leading session awaits the completion of the members and does not proceed prematurely, with some guardrails around it.
Ok, tried to port over to plugin and I have initial version here. I am still not convinced that this should live in a plugin. To fully leverage the capabilities of the harness, this needs to be native.
@CKGrafico Thank you so much for sharing opencode-ensemble and your own project opencode onboard !
@panwar-stack Thank you very much for going the extra mile of releasing a plugin version !
I read the limitations you describe, several points were indeed raised by others and there are some such changes implemented in shuvcode to make subagents states more explorable in the TUI.
Unfortunately i don't expect opencode's team to settle on any specific implementation of agents teams given how orchestration is a hot topic right now and it's chrnging very fast, so I guess the more likely route is to just open a PR with exactly what you wrote about the limitations of the current opencode plugins sdk to ask the team to extend it to expose the parts that would be needed for advanced multiagents orchestration systems such as yours. This does not require them to develop or commit to maintaining a whole new orchestration system this way, they just have to provide interfaces to what is already in the codebase, and this will certainly be useful to a lot of future advanced plugins.
If you open such a pr please link here, i will upvote ! And I advise anyone reading this thread and interested in agents teams and the future orchestration systems beyond to massively upvote too !
@CKGrafico Thank you so much for sharing opencode-ensemble and your own project opencode onboard !
@panwar-stack Thank you very much for going the extra mile of releasing a plugin version !
I read the limitations you describe, several points were indeed raised by others and there are some such changes implemented in shuvcode to make subagents states more explorable in the TUI.
Unfortunately i don't expect opencode's team to settle on any specific implementation of agents teams given how orchestration is a hot topic right now and it's chrnging very fast, so I guess the more likely route is to just open a PR with exactly what you wrote about the limitations of the current opencode plugins sdk to ask the team to extend it to expose the parts that would be needed for advanced multiagents orchestration systems such as yours. This does not require them to develop or commit to maintaining a whole new orchestration system this way, they just have to provide interfaces to what is already in the codebase, and this will certainly be useful to a lot of future advanced plugins.
If you open such a pr please link here, i will upvote ! And I advise anyone reading this thread and interested in agents teams and the future orchestration systems beyond to massively upvote too !
Update on my side, nowadays it is really really easy to have multiple agents that use different skills spawned and working on our features, also sharing their memory. I'm using the native way in ckgrafico/opencode-onboard with good results.
Activity
github-actions commentedon Feb 25, 2026
This issue might be a duplicate of existing issues. Please check:
These issues already discuss the agent-teams feature. Consider adding your thoughts or upvoting there instead of opening a new issue.
OS-Lihua commentedon Feb 26, 2026
Which is the core issue regarding adding Agent Team mode?
boxiaolanya2008 commentedon Feb 26, 2026
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 commentedon Mar 2, 2026
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 commentedon Mar 5, 2026
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 commentedon Mar 8, 2026
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:
archichen commentedon Mar 8, 2026
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
CKGrafico commentedon Apr 19, 2026
Sill trying to get if this is happening or not, I've seen some PRs
lrq3000 commentedon Apr 19, 2026
There are implementations by others already, although I did not yet try them, here is a raw extract from my notes :
allarac commentedon Apr 26, 2026
Teams is a fun new development, but what is an actual problem forcing me to switch to some other framework is the complete inability to spawn multiple subagents at the same time. I sadly cannot wait any longer. It would be a minor change in prompt.ts to allow that.
lrq3000 commentedon Apr 26, 2026
@allarac I thought OpenCode could spawn multiple parallel subagents? At least OpenCode Server appears to allow that. What would need to change in prompt.ts ? If you make a PR this may get merged soon enough.
1 remaining item
allarac commentedon Apr 29, 2026
What are you talking about. It cannot.
panwar-stack commentedon May 2, 2026
Proposal for for adding native support for Agent teams docs: add agent teams proposal
Reference implementation: add agent team orchestration with DB-backed team service, team tools, and HTTP API
lrq3000 commentedon May 3, 2026
I based my assumptions on the fact Kilo Code was rewritten to use OpenCode Server mainly to support parallel tools calls and parallel subagents delegation. So it seems that OpenCode Server is not a blocking point and can support parallel subagents delegation with little changes, if it's not already.
When I use OpenCode it also seems to me it can call and wait for multiple subagents in parallel (using oh-my-openagent), or am I mistaken ?
panwar-stack commentedon May 3, 2026
Example of prompt, conversation and how complex games can be built using agent teams
https://github.com/panwar-stack/shake-rattle-and-bang/tree/main/agent-teams-examples esp this one https://github.com/panwar-stack/shake-rattle-and-bang/tree/main/agent-teams-examples/html5-game
As again, agent teams in below proposal+repo are first-class, persistent multi-agent coordination
Proposal for for adding native support for Agent teams docs: add agent teams proposal
Reference implementation: add agent team orchestration with DB-backed team service, team tools, and HTTP API
lrq3000 commentedon May 3, 2026
@panwar-stack Thank you very much for sharing your work! This is very impressive indeed! I saw the HTML5 game prompt, log and result, it is incredible that you got such a complex fully functional game in one go! You confirm that you inputted just the one prompt you uploaded without any followup?
For those interested, all the relevant commits are in this branch: https://github.com/panwar-stack/opencode/commits/feature/teams-orchestrations/
And yes, agents are spawned and working in parallel!
Given the focus of oh-my-openagent, I am not sure agent teams will get merged, or maybe not anytime soon. Maybe you could consider releasing it as a separate plugin for opencode? I for sure will install and try it (but I will anyway clone your repo to try whenever I have some spare time!).
Incredible work!
CKGrafico commentedon May 3, 2026
I've finally moved to Ensemble, is a nice approach https://github.com/hueyexe/opencode-ensemble and I've also created a simple CLI called opencode onboard that helps developers (mostyl to myself) to add Opencode + Openspec + Ensemble + some other small configurations to brownfield that have no ai configuration installed and need to follow a structure without adding opinionated skills about development https://github.com/CKGrafico/opencode-onboard
panwar-stack commentedon May 3, 2026
@lrq3000 yes, that was all the result of just that one prompt. To be honest, I was not expecting it to work so well on the first go itself. I did have to tweak the implementation so the leading session awaits the completion of the members and does not proceed prematurely, with some guardrails around it.
panwar-stack commentedon May 3, 2026
I would add that credit go to the model as well. Deepseek v4 is very good with tools calling and following instructions
boxiaolanya2008 commentedon May 4, 2026
yes i think the DeepSeek-V4 very too while it is not good at the Claude in the coding but it cheap and cost effort
panwar-stack commentedon May 6, 2026
Ok, tried to port over to plugin and I have initial version here. I am still not convinced that this should live in a plugin. To fully leverage the capabilities of the harness, this needs to be native.
lrq3000 commentedon May 7, 2026
@CKGrafico Thank you so much for sharing opencode-ensemble and your own project opencode onboard !
@panwar-stack Thank you very much for going the extra mile of releasing a plugin version !
I read the limitations you describe, several points were indeed raised by others and there are some such changes implemented in shuvcode to make subagents states more explorable in the TUI.
Unfortunately i don't expect opencode's team to settle on any specific implementation of agents teams given how orchestration is a hot topic right now and it's chrnging very fast, so I guess the more likely route is to just open a PR with exactly what you wrote about the limitations of the current opencode plugins sdk to ask the team to extend it to expose the parts that would be needed for advanced multiagents orchestration systems such as yours. This does not require them to develop or commit to maintaining a whole new orchestration system this way, they just have to provide interfaces to what is already in the codebase, and this will certainly be useful to a lot of future advanced plugins.
If you open such a pr please link here, i will upvote ! And I advise anyone reading this thread and interested in agents teams and the future orchestration systems beyond to massively upvote too !
rekram1-node commentedon Jun 23, 2026
Automated: closing question issues.
GitHub Issues are for bug reports and feature requests only — questions / support are no longer tracked here.
If this is a bug or feature request, please open a new issue with the appropriate template.
CKGrafico commentedon Jun 23, 2026
Update on my side, nowadays it is really really easy to have multiple agents that use different skills spawned and working on our features, also sharing their memory. I'm using the native way in ckgrafico/opencode-onboard with good results.