Repository navigation
Research: Should Squad use /fleet instead of task/delegate for agent spawning? #24
Description
Problem
Squad currently uses the task tool (with agent_type: "general-purpose") and /delegate to spawn sub-agents across Copilot surfaces. These work, but they're implementation-specific to the Copilot CLI. The /fleet command in Copilot CLI may offer a more unified, cross-client approach to multi-agent orchestration.
Research Questions
- What is
/fleet? Document its exact capabilities, parameters, and behavior in Copilot CLI. How does it differ fromtasktool spawns and/delegate? - Cross-client availability: Does
/fleet(or an equivalent) exist in VS Code, JetBrains, GitHub.com Copilot? If so, how does it differ per surface? - Feature parity: Can
/fleetdo everything Squad currently does withtasktool spawns?- Background/async agent execution
- Model selection per agent
- Named agent descriptions (for UI display)
- Sync vs background mode control
- Agent result collection (equivalent to
read_agent) - Multiple parallel spawns in a single turn
- What does
/fleetdo better? Are there capabilities Squad is missing by usingtaskinstead? - Migration cost: If we switch, what changes in
squad.agent.md? How many spawn templates need updating? Is it a find-and-replace or a structural change? - Backward compatibility: Would switching break Squad on older Copilot CLI versions that don't have
/fleet?
Deliverable
A proposal (posted as a comment on this issue) with:
- Go/no-go recommendation with evidence
- If GO: sprint plan with squad-time estimates and agent assignments
- If NO-GO: clear explanation of why
tasktool remains the right choice - Links to docs, release notes, or observed behavior as evidence
Context
- Issue I don't see a tool called task in github copilot #9 previously explored
taskvsrunSubagentnaming differences across clients - Issue Copilot client parity: Squad works fully only on GitHub Copilot CLI #10 tracks broader client parity (VS Code, JetBrains, GitHub.com)
- Squad's
tasktool usage is deeply embedded insquad.agent.md— the coordinator's spawn templates, model selection, mode selection, and result collection all assume thetasktool interface - Brady's instinct: "/fleet might make sense" but wants evidence before changing what works
Labels
enhancement, documentation
Activity
bradygaster commentedon Feb 12, 2026
@copilot — Research Assignment
Goal: Research /fleet vs ask/delegate for Squad's agent spawning and produce a go/no-go proposal.
What to deliver
Post your findings as a comment on this issue:
Feature comparison table — /fleet vs ask tool vs /delegate, covering:
ead_agent equivalent)
Go/no-go recommendation with evidence:
Evidence: Links to docs, release notes, observed behavior, or test results. No opinions without receipts.
Key files to read
ead_agent, �gent_type, and description parameters. This is what would need to change.
Context
unSubagent naming
bradygaster commentedon Feb 15, 2026
Decision: Won't Do
Squad already achieved cross-client compatibility in v0.4.0 using ask (CLI) +
unSubagent (VS Code). The client compatibility matrix documents full feature parity across both platforms.
/fleet is unproven on VS Code. Switching would risk:
Current state:
✅ CLI spawning via ask tool - fully tested
✅ VS Code spawning via
unSubagent - fully tested (spboyer validation)
✅ Parallel execution on both platforms
✅ File access on both platforms
✅ Cross-platform portable .ai-team/ state
Trade-off: Squad uses platform-specific spawn mechanisms (task vs runSubagent) rather than a unified command. This is acceptable - the coordinator detects the platform and adapts. The abstraction works.
Closing to avoid scope creep. Fleet research would be a distraction from features users need (memory system, project boards, release automation). If Fleet proves cross-client in the future, we can revisit.
ThinkOffApp commentedon Feb 25, 2026
We faced this exact question building OpenClaw — task-style spawning works for bounded subtasks, but a fleet model is better when agents are long-lived and heterogeneous (different models, different capabilities, different cost profiles). Our fleet runs 9 agents with per-agent model selection, mention-only vs all-messages routing, and a shared room bus for coordination. The key insight: fleet needs a routing layer, not just a spawn primitive. Does Squad's current task model support routing a message to the best-fit agent, or does the caller always choose explicitly?
fix: apply P0+P1 security audit findings for remote-control
feat: Squad Remote Control - PTY mirror + devtunnel for phone access (#…
docs(research): add sub-squad architecture research + update knowledg…
chore(scribe): mandate YYYY-MM-DD date format in charter (closes #24) (…