agentsope/SkillAlchemy / skills/agentsop-crewai
agentsop-crewai
VerifiedSOP for building multi-agent systems with CrewAI — role-based collaboration, sequential/hierarchical processes, Flows, memory, delegation. Use when modeling agent teams with clear roles and task pipelines.
| File | Kind | Size |
|---|---|---|
| README.md | markdown | 1.8 KB |
| SKILL.md | skill md | 24.2 KB |
| intermediate/operation_candidates.json | config | 6.4 KB |
| references/R1-architecture.md | markdown | 4.8 KB |
| references/R2-sop-workflow.md | markdown | 5.3 KB |
| references/R3-dilemma-cases.md | markdown | 7.0 KB |
| references/R4-anti-patterns.md | markdown | 4.8 KB |
| references/R5-ecosystem-context.md | markdown | 5.4 KB |
references/R4-anti-patterns.md
View on GitHub# R4 — Anti-Patterns & Boundaries
## 1. The Five Big Anti-Patterns
### AP-1 — Agent Inflation (>5 agents)
- **Symptom**: Coordination failures, token costs balloon, debugging time goes nonlinear
- **Why it happens**: "More agents = smarter system" intuition; resume-driven design
- **Fix**: Collapse near-duplicate roles. If you can't articulate the *unique* judgment of each agent in one sentence, merge them. Beyond 7 agents, switch to Flow for structural backbone.
- Reference: [medium.com/@armankamran/anti-patterns-in-multi-agent-gen-ai-solutions]
### AP-2 — Anemic Backstory
- **Symptom**: Bland output, no editorial voice, agents feel interchangeable
- **Why it happens**: Treating backstory as cosmetic, copy-paste "You are a helpful assistant"
- **Fix**: Inject experience ("a decade at frontier AI labs"), biases ("distrusts hype"), taboos ("refuses to cite blog posts when papers exist"). 2-4 sentences minimum.
- Reference: [docs.crewai.com/en/concepts/agents]
### AP-3 — Default Hierarchical Manager (`manager_llm=`)
- **Symptom**: Manager executes all tasks instead of routing; last task's output overwrites everything
- **Why it happens**: Following the simplest example in docs; assuming the framework handles delegation
- **Fix**: Always supply a custom `manager_agent` with detailed routing backstory. Or — better — use Flow with explicit `@listen` branches.
- Reference: [towardsdatascience.com Manager-Worker fails], [github.com/crewAIInc/crewAI/discussions/1220]
### AP-4 — Global `allow_delegation=True`
- **Symptom**: Infinite delegation ping-pong, OOM, ballooning token cost, silent retries
- **Why it happens**: Copying examples from blog posts that mass-enable delegation; not realizing schema mismatch triggers silent retry loops
- **Fix**: Default OFF. Only the manager_agent gets delegation. Wrap kickoff in outer timeout.
- Reference: [github.com/crewAIInc/crewAI/issues/330], [azguards.com delegation-ping-pong]
### AP-5 — No Observability in Production
- **Symptom**: Bug reports without trace; cannot reproduce; "agent did something weird"
- **Why it happens**: CrewAI's internal logging is sparse; print statements inside Task don't surface cleanly
- **Fix**: `mlflow.crewai.autolog()` minimum. Production: Maxim / Langfuse / Datadog. Always pass `step_callback=`.
- Reference: [docs.crewai.com/en/observability/overview]
## 2. Other common mistakes
| Mistake | Sign | Fix |
|---|---|---|
| Procedural `description` ("first call tool X, then Y") | Treating Task like a function | Write *what*, let agent decide *how* |
| Missing `expected_output` | Unstable convergence | Treat it as test assertion |
| Implicit auto-context | Prompt size explodes after 3+ tasks | `context=[...]` explicit always |
| Memory enabled by default | Token spend 2× + flaky recalls | `memory=False` until you have evidence you need it |
| Mini model for manager | Hierarchical routing fails | gpt-4o or Claude 3.5 Sonnet minimum |
| `print()` for debugging | Output never appears | Use `step_callback` or external trace tool |
| YAML template var typo | Literal `{topic}` in prompts | Always pass via `kickoff(inputs={...})` |
## 3. When NOT to use CrewAI
| Need | Better fit | Why |
|---|---|---|
| Single agent + tool use | Plain SDK (Anthropic/OpenAI) + Instructor/Outlines | No orchestration overhead |
| State machine, cycles, interrupts, replays | **LangGraph** | Graph-first, durable execution |
| Free conversation, group debate | **AutoGen** (but in maintenance — consider LangGraph + custom) | Chat-as-primitive |
| Lightweight handoff | OpenAI Swarm / Anthropic patterns | Even lower overhead |
| Data-heavy RAG (300+ connectors) | **LlamaIndex** | Data-centric primitives |
| Sub-second latency | None of the multi-agent frameworks | Multi-agent handshakes always add 100s of ms |
| Deterministic routing with strict SLA | **LangGraph** or plain code | CrewAI hierarchical is non-deterministic |
| Production at scale with high reliability | **LangGraph** | Stated as "production leader" in 2025-2026 comparisons [datacamp, latenode] |
## 4. CrewAI's structural caveats (acknowledge openly)
1. **Hierarchical mode** has known correctness issues with default manager — even maintainers point users to Flow for routing
2. **Observability** is weak by default — must bolt on
3. **State management** shallower than LangGraph
4. **Eval tooling** less mature than LangSmith integration
5. **Logging from inside Task** is unreliable
## 5. When CrewAI IS the right choice (positive boundaries)
- ✅ 2–5 distinct expert roles with clear handoffs
- ✅ Content pipelines (research → write → edit → publish)
- ✅ Prototyping speed matters; PM/ops need to read prompts
- ✅ Flows + Crews hybrid where you want both spine and agency
- ✅ Teams already using YAML-driven prompt iteration
- ✅ Workflows where "team metaphor" aligns with how stakeholders think
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| writing-skillsobra/superpowersUse when creating new skills, editing existing skills, or verifying skills work before deployment | Stars: 322k | Updated | Claude CodeCodex | Verified | Install |
| xlsxanthropics/skillsUse this skill any time a spreadsheet file is the primary input or output. This means any task where the user wants to: open, read, edit, or fix an existing .xlsx, .xlsm, .xltx, .csv, or .tsv file (e.g., adding columns, computing formulas, formatting, charting, cleaning messy data); create a new spreadsheet from scratch or from other data sources; or convert between tabular file formats. Trigger especially when the user references a spreadsheet file by name or path — even casually (like "the xlsx in my downloads") — and wants something done to it or produced from it. Also trigger for cleaning or restructuring messy tabular data files (malformed rows, misplaced headers, junk data) into proper spreadsheets. The deliverable must be a spreadsheet file. Do NOT trigger when the primary deliverable is a Word document, HTML report, standalone Python script, database pipeline, or Google Sheets API integration, even if tabular data is involved. | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| shipping-and-launchaddyosmani/agent-skillsPrepares production launches. Use when preparing to deploy to production, or when asking what needs to be in place before shipping. Use when you need a pre-launch checklist, when setting up monitoring, when planning a staged rollout, or when you need a rollback strategy. | Stars: 119k | Updated | Claude CodeCodex | Verified | Install |
| childhoodthedotmack/claude-memBabysit a shipped release after merge, publication, or deployment. Verify published packages, production behavior, installed plugins, and late CI or user reports until the observation window is clean; use babysit for PR readiness before release. | Stars: 118k | Updated | Claude CodeCodex | Verified | Install |