Multi-Agent Crew Design

Why#

The current 4-agent crew evolved organically during the first week. Roles overlap, Robbie carries too much load, and there's no dedicated cost monitoring or code review. This project designs and builds the right team.

Design Principles#

1. Match model to job difficulty — DeepSeek Flash for reasoning/active work. Local Qwen 7B for routine/cron. No waste.
2. Privacy-sensitive work stays local — credentials, personal data, local-only analysis never touch cloud.
3. Coordinate through workspace files — task files, memory, team briefs. No exotic protocols.
4. Eliminate overlap — every agent has a clear domain. No two agents doing the same thing.
5. Cost matters — DeepSeek cloud ($) and local (free). No Anthropic unless forced.

Crew Composition#

🧐 Robbie — Chief of Staff#

• Model: DeepSeek Flash (cloud)
• Good at: Strategy, architecture, complex multi-step reasoning, cross-agent orchestration, judgment calls, anything requiring tool chains across multiple systems
• NOT good at: High-volume batch work (too expensive), rote single-step tasks (waste of capacity), things he should delegate
• Job: Maxime's primary interface. Makes decisions, distributes work, handles the hard stuff. Delegates everything else.
• Cost: $ (mid)

🐯 Chris — Build Engineer#

• Model: DeepSeek Flash (cloud)
• Good at: Multi-file code changes, scripts, automation, infra config, complex build tool chains, following a spec precisely
• NOT good at: Creative/strategic decisions (needs clear specs), documentation writing (too verbose), judgment calls on what to build
• Job: Mission Control and workspace code. Frontend, backend, scripts, cron job creation. Executes well-defined build tasks.
• Cost: $ (mid)

🦑 Debbie — Operations Worker#

• Model: Qwen 2.5 7B (local — FREE, privacy-safe)
• Good at: Fast batch processing, classification, summarization, running cron jobs, health checks, queue draining, things that run in isolation without history
• NOT good at: Anything needing tool calling (Qwen flakes on multi-step tools), complex reasoning or multi-turn conversations, writing production code
• Job: High-volume ops. Cron tasks, health probes, classification/triage, routine summaries. Writing daily activity logs.
• Cost: Free

🌺 Kelly — Documentation Keeper#

• Model: DeepSeek Flash (cloud)
• Good at: Reading code and documentation, finding inconsistencies, updating READMEs/admin files, cross-referencing multiple files for consistency
• NOT good at: Writing new code (she can read it but not ship it), making architectural decisions, anything outside documentation domain
• Job: Daily doc sweep at 03:00 SGT. Audits every project's docs for accuracy. Reviews new agents' identity files. Fixes what's stale. Updates CREW-MANIFEST.
• Cost: $ (mid)

📊 Oscar — Cost Monitor#

• Model: Qwen 2.5 7B (local — FREE, privacy-safe)
• Good at: Reading numeric data, comparing time series, flagging anomalies, generating simple reports
• NOT good at: Complex reasoning (Qwen), inference across multiple domains, estimating future costs with uncertain inputs
• Job: Reads usage cost cache daily at 08:00 SGT. Compares spend week-over-week. Forecasts month-end cost. Alerts if spending spikes.
• Cost: Free

🔍 Vera — Code Reviewer#

• Model: DeepSeek Flash (cloud)
• Good at: Reading diffs, finding bugs, style issues, performance problems, security concerns. Thorough and pedantic.
• NOT good at: Writing new features (reviewer, not builder), understanding business context without a spec, fixing what she finds
• Job: Reviews commits on mission-control repo every 2h. Posts findings to Discord #chris. Weekly deep-dive on quality patterns.
• Cost: $ (mid)

❌ Nora — (REMOVED)#

Inbox triage wasn't needed — Maxime has no email/mention workflow that needs an agent.

Estimated Costs#

• 3 agents on DeepSeek Flash × modest daily usage ≈ ~$1-3/month total
• 2 agents on local Qwen 7B: $0 (electricity only)
• Total crew cost: ~$1-3/month

Privacy Boundary#

• LOCAL ONLY (Qwen 7B): Cost data (Oscar), operational logs (Debbie). Nothing sensitive leaves the machine.
• CLOUD OK (DeepSeek): Code, docs, task management, public web data, general reasoning.

Task dependency order#

1. t-072 — Robbie SOUL.md rewrite (foundation — clear delegation rules for the whole crew)
2. t-074 — Oscar setup (simple, local, proves new-agent pipeline)
3. t-075 — Vera setup (needs post-Oscar momentum)
4. t-076 — Debbie + Kelly scope refinement (after new agents exist)
5. t-077 — Kelly updates CREW-MANIFEST (last — depends on all above)