/team - Parallel Multi-Agent Execution

In the previous article, /run built the homepage and navigation from the design spec. Now you have three pages left - About, Portfolio, and Blog listing - and they're all independent. None of them depend on each other's output. This is exactly what /team is for.

/team "Build the remaining three pages for the portfolio site:
1. About page - bio, skills, and a contact form. Pull from the /designer spec at design/about-spec.md
2. Portfolio page - project grid with filtering by technology. Pull from design/portfolio-spec.md
3. Blog listing page - article cards with tag filtering and pagination. Pull from design/blog-spec.md

Each page should match the design system established in the homepage."

Once you submit, /team starts analysis:

[team] Analyzing task list...
[team] Identified 3 independent tasks → assigning to Wave 1
[team] Spawning agents...

[wave-1] about-agent    → Starting: About page
[wave-1] portfolio-agent → Starting: Portfolio page
[wave-1] blog-agent     → Starting: Blog listing page

All three agents are running simultaneously. You'll see interleaved progress:

[about-agent]     Creating components/About/Bio.tsx
[portfolio-agent] Creating components/Portfolio/ProjectGrid.tsx
[blog-agent]      Creating components/Blog/ArticleCard.tsx
[about-agent]     Creating components/About/ContactForm.tsx
[portfolio-agent] Implementing tag filter logic
[blog-agent]      Adding pagination controls

When all three finish, the quality gate runs:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] about-agent    ✓ PASS - About page complete, form validation included
[quality-gate] portfolio-agent ✓ PASS - Portfolio grid renders, filter works
[quality-gate] blog-agent     ✓ PASS - Listing page complete, pagination functional

[team] Wave 1 complete. All 3 tasks passed quality gate.
[team] No wave 2 required.

Summary:
  ✓ About page        components/About/
  ✓ Portfolio page    components/Portfolio/
  ✓ Blog listing      components/Blog/

Three pages that would have taken three sequential /run invocations - and the wait for each to finish before starting the next - are done in the time it takes to build one. The first time I ran /team on a batch like this, I kept checking the terminal because I didn't believe all three were actually running. They were.

In the previous article, /run produced the first blog post from the content strategy spec. Now you need four more posts to fill out the first month's content calendar. They're all on different topics, they don't reference each other, and they all follow the same brief. This is textbook /team territory.

/team "Write four blog posts for the first month's content calendar.
Each post should follow the voice and format guidelines in content/style-guide.md.

Posts to write:
1. '5 Things I Learned Migrating to Kubernetes' - audience: mid-level DevOps engineers
2. 'Why Your Monitoring Setup Is Lying to You' - audience: SREs and platform engineers
3. 'The Case for Boring Technology' - audience: engineering leads and CTOs
4. 'Git Workflows for Small Teams' - audience: developers at startups

Target length: 1200–1500 words each. SEO metadata included."

Four agents spin up immediately:

[team] 4 independent tasks identified → Wave 1
[team] Spawning agents...

[wave-1] post-1-agent → Starting: Kubernetes post
[wave-1] post-2-agent → Starting: Monitoring post
[wave-1] post-3-agent → Starting: Boring technology post
[wave-1] post-4-agent → Starting: Git workflows post

Each agent works independently, writing its post to match the style guide. No agent is waiting for another. When all four finish:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] post-1-agent ✓ PASS - 1,340 words, metadata included
[quality-gate] post-2-agent ✓ PASS - 1,280 words, metadata included
[quality-gate] post-3-agent ✓ PASS - 1,420 words, metadata included
[quality-gate] post-4-agent ✓ PASS - 1,190 words, needs minor length adjustment

[team] Wave 1 complete. 4/4 tasks passed quality gate.

Outputs:
  ✓ content/posts/kubernetes-lessons.md
  ✓ content/posts/monitoring-lies.md
  ✓ content/posts/boring-technology.md
  ✓ content/posts/git-workflows.md

A month's content calendar, written in parallel. The quality gate flagged the fourth post as slightly short - not a hard failure, just a note in the report. You can decide whether to follow up with /run "Expand git-workflows.md to 1200 words" or leave it.

In the /run article, Maya built the complete materials package for Unit 1 of her redesigned US History course -- project brief, 15 lesson plans, source documents, rubric, and parent letter. Now she needs Units 2 through 6, and none of them depend on each other. They all follow the same structure, just different topics.

/team "Build materials packages for Units 2-6, following the same structure as
the Unit 1 package. Each unit needs: project brief, 15 daily lesson plans,
4 primary source documents with questions, rubric, parent update letter.

Unit 2: Westward Expansion - Whose Manifest Destiny?
Unit 3: Civil War - Letters Home (epistolary project)
Unit 4: Reconstruction - Unfinished Revolution
Unit 5: Industrialization - The Price of Progress
Unit 6: Progressive Era - Reform Then and Now"

Five agents, one per unit, all in Wave 1:

[team] Analyzing task list...
[team] Identified 5 independent tasks → assigning to Wave 1
[team] Spawning agents...

[wave-1] unit-2-agent → Starting: Westward Expansion
[wave-1] unit-3-agent → Starting: Civil War
[wave-1] unit-4-agent → Starting: Reconstruction
[wave-1] unit-5-agent → Starting: Industrialization
[wave-1] unit-6-agent → Starting: Progressive Era

All five run simultaneously. The interleaved output shows each agent working through the same structure on different material:

[unit-2-agent] Creating project brief: land grant analysis + indigenous perspectives
[unit-3-agent] Creating project brief: epistolary project framing
[unit-5-agent] Creating project brief: factory conditions investigation
[unit-4-agent] Creating project brief: constitutional amendment analysis
[unit-6-agent] Creating project brief: then-vs-now reform comparison
[unit-3-agent] Adapting lesson plan structure for letter-writing format
[unit-2-agent] Sourcing primary documents: Homestead Act, treaty texts
[unit-5-agent] Sourcing primary documents: factory inspection reports, photographs
[unit-4-agent] Building rubric: amendment analysis criteria
[unit-6-agent] Sourcing primary documents: muckraker excerpts, labor law texts

Unit 3 adapts the lesson plan format rather than forcing the template -- an epistolary project needs writing workshops, not just research days. When all five finish:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] unit-2-agent ✓ PASS - 15 plans, 4 sources, rubric aligned
[quality-gate] unit-3-agent ✓ PASS - adapted structure for epistolary format
[quality-gate] unit-4-agent ✗ FLAG - rubric uses different criteria weighting
              than Units 1-3 (equal weight across categories vs. tiered)
[quality-gate] unit-5-agent ✓ PASS - all materials complete
[quality-gate] unit-6-agent ✓ PASS - contemporary parallels well-sourced

[team] Wave 1 complete. 4/5 passed. 1 flagged for revision.

Outputs:
  ✓ Unit 2: Westward Expansion     curriculum/unit-2/
  ✓ Unit 3: Civil War              curriculum/unit-3/
  ⚠ Unit 4: Reconstruction         curriculum/unit-4/ (rubric inconsistency)
  ✓ Unit 5: Industrialization       curriculum/unit-5/
  ✓ Unit 6: Progressive Era         curriculum/unit-6/

The quality gate caught that Unit 4's rubric weighted all criteria equally while Units 1-3 used tiered weighting (content accuracy weighted higher than presentation). That kind of inconsistency is invisible when you're reviewing one unit at a time -- students would notice it when their grading criteria suddenly shift mid-semester. A quick /run "Revise Unit 4 rubric to match the tiered weighting from Units 1-3" fixes it.

75 lesson plans, 20 source documents, 5 project briefs, 5 rubrics, and 5 parent letters. The semester's curriculum, built in the time it took to produce one unit.

In the /run article, David built the operations playbook for his second restaurant location -- opening checklists, prep schedules, recipe cards scaled for the smaller kitchen. Now he needs three independent workstreams done simultaneously: staffing, training, and the customer-facing menu.

/team "Three workstreams for the Decatur location:
1. Staffing plan: 3 job descriptions (line cook, prep cook, server), interview
   question guides for each, and an onboarding checklist.
2. Training manual: kitchen procedures, recipe execution standards, customer
   service guidelines. Based on the ops playbook we just built.
3. Menu design: 12-item menu with descriptions, pricing (15% above Midtown to
   account for Decatur rent), allergen indicators, and suggested pairings."

Three agents, three different document types:

[team] Analyzing task list...
[team] Identified 3 independent tasks → assigning to Wave 1
[team] Spawning agents...

[wave-1] staffing-agent  → Starting: Staffing plan
[wave-1] training-agent  → Starting: Training manual
[wave-1] menu-agent      → Starting: Menu design

Each agent works in its own domain -- HR, operations, customer-facing:

[staffing-agent]  Drafting line cook job description with salary range
[menu-agent]      Writing descriptions for jollof rice, suya, egusi soup
[training-agent]  Documenting kitchen open procedures from ops playbook
[staffing-agent]  Building practical skills test for prep cook interviews
[menu-agent]      Calculating Decatur pricing: 15% adjustment per item
[training-agent]  Writing recipe execution standards for 12 launch items
[staffing-agent]  Creating 90-day onboarding milestone checklist
[menu-agent]      Adding allergen indicators and pairing suggestions
[training-agent]  Documenting sauce station prep workflow

When all three finish:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] staffing-agent  ✓ PASS - 3 roles, interview guides, onboarding plan
[quality-gate] training-agent  ✗ FLAG - references "sauce station" prep procedure,
              but Decatur kitchen has no dedicated sauce station (smaller layout)
[quality-gate] menu-agent      ✓ PASS - 12 items, pricing verified, allergens marked

[team] Wave 1 complete. 2/3 passed. 1 flagged for revision.

Outputs:
  ✓ Staffing plan      staffing/decatur/
  ⚠ Training manual    training/decatur/ (equipment mismatch)
  ✓ Menu design        menu/decatur-launch/

The training manual was built from the ops playbook, which was built from the Midtown kitchen layout. The Midtown kitchen has a dedicated sauce station. The Decatur kitchen doesn't -- it's a smaller space where sauces are prepped at a shared station. A training manual that tells new cooks to "report to the sauce station at 8am" would confuse every person on their first day. The quality gate caught a mismatch between the plan and the physical space -- exactly the kind of cross-document consistency error that causes real problems during a restaurant opening.

In the /run article, Sadie built the press kit and rewrote her Steam page copy for Meridian. Now she's in weeks 7-5 of the launch countdown, and three workstreams need to happen at once: a trailer script she'll capture footage for, community updates to build anticipation, and personalized press outreach to 30 journalists.

/team "Three launch prep workstreams:
1. Trailer script: 90-second narrative trailer for Meridian. Shot list with
   in-game scenes to capture, voiceover script, music timing notes.
2. Community posts: 6 weekly Steam community hub updates counting down to launch.
   Tone: personal, behind-the-scenes, building anticipation without being salesy.
3. Press outreach: 30 personalized pitch emails for games journalists and
   streamers who cover narrative/puzzle games. Each email references something
   specific the journalist has covered. Pull from the press list in the strategy."

All three are independent -- the trailer doesn't need the press emails, the community posts don't need the trailer:

[team] Analyzing task list...
[team] Identified 3 independent tasks → assigning to Wave 1
[team] Spawning agents...

[wave-1] trailer-agent   → Starting: Trailer script
[wave-1] community-agent → Starting: Community posts
[wave-1] press-agent     → Starting: Press outreach (30 emails)

The press agent has the heaviest workload -- 30 individually tailored pitches:

[trailer-agent]   Structuring narrative arc: quiet gameplay → puzzle reveal → emotional hook
[community-agent] Writing post 1: "What I've been working on for 2 years"
[press-agent]     Personalizing pitch 1/30: referencing journalist's Return of the Obra Dinn review
[trailer-agent]   Building shot list: 14 in-game scenes with capture instructions
[community-agent] Writing post 4: "Release date announcement"
[press-agent]     Personalizing pitch 12/30: referencing streamer's narrative puzzle playlist
[trailer-agent]   Writing voiceover script with music timing cues
[press-agent]     Personalizing pitch 23/30: referencing journalist's indie spotlight column
[community-agent] Writing post 6: "Launch day — here we go"
[press-agent]     Personalizing pitch 30/30: referencing streamer's Baba Is You deep dive

Quality gate reviews all three:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] trailer-agent   ✓ PASS - 90-sec script, 14 shots, music cues included
[quality-gate] community-agent ✓ PASS - 6 posts, escalating anticipation arc
[quality-gate] press-agent     ✗ FLAG - pitch #17 references journalist who hasn't
              published in 8 months (possibly inactive or changed outlets)

[team] Wave 1 complete. 2/3 passed. 1 flagged for revision.

Outputs:
  ✓ Trailer script      launch/trailer-script.md
  ✓ Community posts      launch/community/
  ⚠ Press outreach       launch/press/ (1 of 30 emails flagged)

Writing 30 individually personalized emails is the kind of work that's genuinely tedious to do well and genuinely terrible when done as obvious mail-merge. "I noticed you covered [GAME] -- Meridian shares [SIMILARITY]" stops working the instant a journalist recognizes the pattern. The quality gate flagging an inactive journalist is a small catch, but emailing someone who left their outlet 8 months ago signals that you didn't actually research them. One bad pitch in a batch of 30 is fine to fix with a quick follow-up.

In the /run article, Rosa built the bilingual vendor RFP package and the guest communication suite. Vendors are confirmed. Now she needs three documents that together define the wedding day: the ceremony script, the reception timeline, and the coordination document she'll carry in her hand on the day itself.

/team "Three event documents:
1. Ceremony script: non-religious, bilingual English/Spanish, 25 minutes.
   Includes personal vows section, ring exchange, unity ceremony (sand
   blending, representing two families). Tone: warm, not formal.
2. Reception timeline: cocktail hour through last dance. Minute-by-minute from
   5:30pm to 11:00pm. Includes toasts, first dance, dinner service, cake,
   bouquet toss, and DJ set list transitions.
3. Day-of coordination document: every vendor's arrival time, setup location,
   contact info, meal provisions, and their specific cues. This is the document
   Rosa will carry on the day."

Three documents, all independent, all Wave 1:

[team] Analyzing task list...
[team] Identified 3 independent tasks → assigning to Wave 1
[team] Spawning agents...

[wave-1] ceremony-agent    → Starting: Ceremony script
[wave-1] reception-agent   → Starting: Reception timeline
[wave-1] coordination-agent → Starting: Day-of coordination doc

Each agent handles a different slice of the same event:

[ceremony-agent]     Writing bilingual ring exchange passage
[reception-agent]    Mapping cocktail hour → dinner service transition
[coordination-agent] Compiling vendor arrival times and setup locations
[ceremony-agent]     Structuring sand ceremony with bilingual narration
[reception-agent]    Adding 10-minute buffers between events (toasts run long)
[coordination-agent] Listing vendor meal provisions and dietary restrictions
[ceremony-agent]     Finalizing pronouncement in English and Spanish
[reception-agent]    Mapping DJ set transitions: dinner → dancing
[coordination-agent] Writing vendor-specific cue sheet for each arrival

Quality gate checks all three against each other:

[quality-gate] Reviewing wave 1 outputs...
[quality-gate] ceremony-agent     ✓ PASS - 25 min script, bilingual, warm tone
[quality-gate] reception-agent    ✓ PASS - full timeline with buffers
[quality-gate] coordination-agent ✗ FLAG - florist arrival listed at 2:00pm,
              but ceremony arch arrangement requires 3 hours setup.
              Ceremony starts at 4:30pm. Florist needs to arrive by 1:00pm.

[team] Wave 1 complete. 2/3 passed. 1 flagged for revision.

Outputs:
  ✓ Ceremony script         wedding/ceremony-script.md
  ✓ Reception timeline       wedding/reception-timeline.md
  ⚠ Day-of coordination     wedding/day-of-coord.md (arrival time conflict)

The florist arrives at 2pm. The ceremony starts at 4:30. The arch arrangement takes 3 hours. That's a ceremony with no arch, or a ceremony that starts 30 minutes late while everyone watches a florist work. It's a small detail in a large document -- the kind of thing you catch at 11pm the night before and scramble to fix, or the kind of thing a quality gate catches three weeks early. The coordination document gets one line changed. The wedding stays on schedule.

In the /run article, Jordan wrote the full production script for Episode 1 of Forgotten Picket Lines. Now they need to stay ahead of the pipeline: episodes 2, 3, and 4 each need a research package first, then a script built from that research. The scripts depend on the research -- you can't write a historically accurate narrative without the sources.

This is the first time /team runs a two-wave execution. Wave 1 produces research. Wave 2 uses that research to write scripts.

/team "Build research packages and draft scripts for episodes 2-4:
1. Ep 2: 'The Ludlow Massacre' (1914 Colorado coal miners) - 30 min narrative
2. Ep 3: 'Memphis, 1968' (sanitation workers' strike, MLK connection) - 32 min
3. Ep 4: 'The Flint Sit-Down' (1936-37 GM strike) - 28 min

Each episode needs: research package (10+ sources, timeline, key figures,
narrative arc suggestion) and full production script in the same format as
Episode 1."

/team detects the dependency and splits the work into two waves:

[team] Analyzing task list...
[team] Detected dependency: scripts require research packages as input
[team] Wave 1: 3 research tasks (parallel)
[team] Wave 2: 3 script tasks (parallel, depends on Wave 1)
[team] Spawning Wave 1 agents...

[wave-1] research-2-agent → Starting: Ludlow Massacre research
[wave-1] research-3-agent → Starting: Memphis 1968 research
[wave-1] research-4-agent → Starting: Flint Sit-Down research

Three research packages built simultaneously:

[research-2-agent] Compiling sources: Colorado coal mine records, tent colony accounts
[research-3-agent] Compiling sources: sanitation workers' organizing documents, MLK speeches
[research-4-agent] Compiling sources: GM strike records, UAW founding documents
[research-2-agent] Building timeline: April 1914 events at Ludlow
[research-3-agent] Mapping key figures: T.O. Jones, Jerry Wurf, Martin Luther King Jr.
[research-4-agent] Building timeline: December 1936 through February 1937
[research-2-agent] Suggesting narrative arc: open from child's perspective in tent colony
[research-3-agent] Suggesting narrative arc: trace the "I Am a Man" signs to their origin
[research-4-agent] Suggesting narrative arc: the moment the workers locked the doors

Wave 1 finishes. Quality gate runs before any scripts are written:

[quality-gate] Reviewing Wave 1 outputs...
[quality-gate] research-2-agent ✓ PASS - 14 sources, timeline complete
[quality-gate] research-3-agent ✓ PASS - 12 sources, but NOTE: episode intersects
              with MLK's assassination (April 4, 1968). Recommend consulting
              Memphis community voices before scripting. Sensitivity flag
              added to Wave 2 brief.
[quality-gate] research-4-agent ✓ PASS - 11 sources, timeline complete

[team] Wave 1 complete. 3/3 passed. 1 sensitivity note forwarded to Wave 2.
[team] Spawning Wave 2 agents...

[wave-2] script-2-agent → Starting: Ludlow script (using research package)
[wave-2] script-3-agent → Starting: Memphis script (using research + sensitivity note)
[wave-2] script-4-agent → Starting: Flint script (using research package)

Wave 2 agents write from the research, not from scratch:

[script-2-agent] Opening with tent colony scene — child's perspective per research arc
[script-3-agent] Structuring narrative around "I Am a Man" — noting sensitivity boundaries
[script-4-agent] Opening with the factory floor moment — workers locking the doors
[script-2-agent] Integrating primary source: Mary Petrucci's account of the fire
[script-3-agent] Framing MLK connection through workers' perspective, not assassination
[script-4-agent] Writing interview segment placeholder: labor historian on sit-down legacy
[script-2-agent] Adding production notes: [PACE: hold 4 seconds after the fire description]
[script-3-agent] Adding production notes: [MUSIC: restraint here — dignity, not tragedy]
[script-4-agent] Finalizing at 28 minutes, within target

Wave 2 quality gate:

[quality-gate] Reviewing Wave 2 outputs...
[quality-gate] script-2-agent ✓ PASS - 30 min, sources cited, production notes included
[quality-gate] script-3-agent ✓ PASS - sensitivity handled well, MLK framed through
              workers' agency rather than assassination narrative
[quality-gate] script-4-agent ✗ FLAG - script runs 38 minutes (target: 28 min)

[team] Wave 2 complete. 2/3 passed. 1 flagged for trimming.

Summary:
  Wave 1 (research):
    ✓ Ep 2 research    episodes/ep02/research.md
    ✓ Ep 3 research    episodes/ep03/research.md
    ✓ Ep 4 research    episodes/ep04/research.md
  Wave 2 (scripts):
    ✓ Ep 2 script      episodes/ep02/script.md
    ✓ Ep 3 script      episodes/ep03/script.md
    ⚠ Ep 4 script      episodes/ep04/script.md (over length target)

Two waves, six agents total, three research packages feeding three scripts. The quality gate between waves did something more meaningful than checking word counts -- it flagged a content sensitivity issue and passed it forward so the script agent could handle it thoughtfully. The Memphis episode frames MLK through the workers' story rather than the assassination narrative, which is both more respectful and a better episode. The Flint script came in long; Jordan can trim it with a quick /run follow-up or decide that 38 minutes is fine for that story.

This is where /team earns its keep over running six separate /run commands. The wave structure means the research is validated before anyone writes a word of script. The sensitivity flag propagates automatically. And the whole thing runs in two parallel batches instead of six sequential tasks.