50 min orchestration, 25 min hands on, the rest together
0:00
Orchestration
0:50
Break
0:55
In the session·generalize one workflow a team already automated
1:20
Homework, close
1:30
End
Where we are
01Core concepts
02Preparing the codebase for agents
03Writing a good spec
04Feedback loops
05Workflow integration and code review
06Claude Code in depth
07TodayOrchestration and subagents
08Claude Code alternatives
One tool, several agents. Everything so far is the precondition.
One agent, one task so far. Today: several at once.
What we cover today
Agent view/batchExploreRoutine
WorktreeSubagent
Agent teamclaude -pPluginMarketplace
Dynamic workflowCross-session messagingOwner
Four ways to run agents in parallel
Subagents
Agent view
Agent teams
Workflows
Who coordinates
Claude, in your session
You
A lead agent
A script
Who they report to
the session
you
each other
the script
Status
built in
research preview
experimental
built in
Three questions decide: who coordinates, do they need to talk, do they touch the same files.
Part one
Working withsubagents
Delegated work in its own context.
Subagent its own context, for one narrow job
When a subagent beats one long session.
Research, search, review across many files: anywhere the result matters more than the path.
Custom subagents
.claude/agents/checkout-reviewer.md
---
name: checkout-reviewer
description: Reviews changes under src/checkout/ against our rules
tools: Read, Grep, Glob, Bash
model: opus
---
Review the diff against docs/standards/checkout.md.
Report only findings that affect correctness or the spec.
/subtask hands a side task to a subagent that inherits the whole conversation.
Part two
Manysessions
Worktrees, agent view, /batch, workflows, teams.
Worktrees in practice
terminal
claude -w voucher-fix # a new session in its own worktree
cat .worktreeinclude # files git ignores, copied into each worktree
.env.local
Finished worktrees need cleaning up. Session 05: a branch for one agent, a worktree each for several.
Agent view
One screen for every background session, and which one needs you.
claude agents. Each session moves into its own worktree before it edits. [VERIFY status: research preview]
/batch
One change split across 5 to 30 subagents, each in its own worktree, each with a pull request.
Dynamic workflow
A script holds the plan: many subagents, findings cross-checked against each other.
Codebase audits, large migrations, research from several angles.
Agent team
A lead agent assigns work to teammates who message each other.
Experimental and off by default. No separate worktrees: split the files between them yourself.
Cross-session messaging
Sessions you run yourself pass findings to each other.
Without a person at the keyboard
Non-interactive
claude -p: one prompt, fixed permissions, in scripts and CI.
CI
GitHub Actions or GitLab CI: @claude on an issue or pull request.
Routines
Scheduled or triggered by an event, running in the cloud.
Part three
Sharingwhat works
From one person's setup to team infrastructure.
The orchestrator is you.
Splitting work into independent tasks with clear specs is the skill. The tools only run what you split.
Upkeep across many systems.
One dispatched session per application, same routine task, watched from agent view.
A working setup is infrastructure once another team can install it.
Plugin
Skills, subagents, hooks and MCP servers in one install.
A team marketplace distributes it. Plugin evals test it against a run without it, in CI.
Shared infrastructure needs an owner.
Versioned, reviewed, and someone who answers when it breaks.
Where orchestration stops paying
Review time grows faster than the time saved
The tasks were not independent after all
Token use multiplies with every parallel session
In the session
Generalize one workflow a team already automated.
Pick the one furthest along · goal: another team can install it today
Example homework, to be tailored to your team
Homework: three options
Cross-team
Install and break
Install what another team built, run it on a real ticket, report what broke. The builders fix it and add a plugin eval.
Team build
One unattended job
claude -p from CI, or a routine, for a routine task. Bring a week of its output.
Experiment
Three tickets in parallel
Three independent tickets from agent view, a worktree each. Wall-clock and review time against one after another.
Next session, the last one: Claude Code alternatives.