- Home
- Skills
- Agents & Orchestration
- Agentic AI Orchestration
More screenshots
Works with the AI Agents you already use
Agentic AI Orchestration
Design, build and test AI agents the sensible way: ReAct loops, plan-and-execute, supervisor graphs, LangGraph and CrewAI, with approval gates, budgets and a test harness you can run on a local Llama model.
$12
Agentic AI Orchestration
Example session with this skill installed
Build me a LangGraph agent that checks three competitors' pricing pages and drafts a summary email. It must ask me before sending anything. Keep it cheap to run on a local Llama model.
- Read your context and instructions
- Compiled the agentic ai orchestration
Summary
Pattern chosen
LangGraph ReAct agent with an approval gate. The task is focused, has two tools, and has one irreversible action (sending email), so a full multi-agent setup would add cost without adding control.
Design
| Decision | Choice | Why |
|---|---|---|
| Pattern | Single ReAct loop | Two tools, one goal |
| Tools | fetch_page, send_email | Fewer than 3 tools suits small local models |
| Risky tool | send_email | Pauses for human approval |
| Budget | 6 steps, 3 retries | Stops loops early |
| Model | Llama via Ollama, temperature 0 | Repeatable runs |
State
class AgentState(TypedDict):
messages: list
competitors: list[str]
summary: str
approved: bool
steps_used: int
Flow
- Fetch each competitor page (a failed fetch is recorded and skipped, not retried forever).
- Draft the summary from the fetched text only.
- Pause for approval. Nothing is sent until you reply "approve".
- Send on approval; otherwise return the draft unchanged.
Tests to run before trusting it
- A page that times out: the agent reports it and continues
- You reject the draft: no email is sent
- Step budget reached: the agent stops with a clear message
Next steps
- Run
python -m pytest assets/tests -qto confirm the skeletons work on your machine - Run
eval-harness.pyon your own model to see real cost and success numbers - Swap the stub search tool for a real one before relying on the results
Connects securely to your tools. The creator never sees your data.
About this skill
Most agent demos work once and fall apart on the second run. This skill helps your AI assistant build agents that behave: they have a clear state, sensible limits, and a way to stop before they run away.
Install it, then ask your agent for a tool-calling agent, a LangGraph workflow, a CrewAI crew or a supervisor setup. It picks a pattern first, explains why, and only then writes code.
What you get
- Pattern-first design. ReAct tool loops, plan-and-execute, supervisor and workers, hierarchical crews, and explicit LangGraph graphs. Your agent chooses one and names it before writing anything.
- Safety built in. Step and token budgets, retry limits, human approval before risky actions, and clear failure messages instead of silent loops.
- Working starter code. Tested skeletons for LangGraph ReAct, LangGraph plan-and-execute (with tool support), CrewAI, and a competitor-pricing example agent.
- A test harness.
eval-harness.pycompares ReAct and plan-and-execute on the same tasks and reports success, LLM calls, tokens and time. It prints every answer so you can check the results yourself. - Local model support. Run everything on Llama or Qwen through Ollama, with notes on what small models can and can't do.
- 16 offline tests that run without an API key.
How we tested it
We ran the harness on two small local models (llama3.2:3b and qwen2.5:3b) on an ordinary Intel i3 laptop, one run per task, temperature 0.
| Model | ReAct passed | Plan-Execute passed | ReAct tokens | Plan-Execute tokens |
|---|---|---|---|---|
| llama3.2:3b | 6 / 7 | 7 / 7 | 452 | 1,184 |
| qwen2.5:3b | 7 / 7 | 4 / 7 | 501 | 2,096 |
The clearest finding is cost: plan-and-execute used about 2.6x to 4.2x the tokens and 2.6x to 3.4x the time of ReAct. Every failed task was a wrong arithmetic answer from a 3B model, not a skill bug. This is a small sample (7 scored tasks, one run each), so treat it as a regression check, not a benchmark. Full details are in references/test-results.md.
See the full results, every raw answer and our test setup: Open test results
Good fit if you
- are building your first serious agent and want a sound structure
- already use LangGraph or CrewAI and want guardrails and tests
- want to try agents on a local model without an API bill
Not a fit if you
- want a hosted agent platform or a no-code builder (this is guidance and starter code for your own agent)
- expect small local models to match large hosted ones on hard reasoning.
Find more skills like this on the Claude Skills Marketplace, the Codex Skills Marketplace and the Gemini Skills Marketplace, or browse Claude Skills for Agents & Orchestration.
We'd rather you know these before buying:
- Tested on two small models only. We ran ReAct and plan-and-execute live on llama3.2:3b and qwen2.5:3b (via Ollama, Windows, one run per task). We haven't benchmarked larger local models or hosted APIs, so run the included harness on yours.
- Small sample. 7 scored tasks per model, one run each. Treat a one-task difference as noise.
- Only two patterns have tested code. ReAct and plan-and-execute have starter code and measured results. Supervisor + workers and hierarchical crews are guidance plus a CrewAI skeleton. We haven't benchmarked them.
- The CrewAI skeleton isn't run end to end. We verified that it builds. A full run needs your own model.
- The example search tool is a stub. It just echoes the query. Swap in a real one before trusting search results.
- Failure checks are simple keyword rules. Replace them with checks that suit your own task.
- Small models make arithmetic mistakes. Every failure in our runs was a wrong number from a 3B model, not a skill bug. Give agents a calculator tool.
- Approval pauses need a checkpointer. The in-memory one resets when the program stops. Use SQLite or Postgres for real resumability.
- Budgets count steps, not money. Add provider-level cost caps in production.
- Frameworks change quickly. Tested with LangGraph 1.2.x and CrewAI 1.15.x. Pin your versions and re-run the tests after upgrading.
More detail on what we tested and what we didn’t:https://alphazolo12200-tech.github.io/agentic-ai-orchestration-test-results/
How to use
- 1
Connect your agent to the Agensi MCP
Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup
- 2
Sign in to Agensi
Your agent shows a sign-in link. Sign in and click Allow.
- 3
Ask for what you need
Your agent loads this skill and uses it on your task. It always asks before buying a skill.
Reviews
No reviews yet
Be one of the first to try it. Every listed skill goes through the trust checks below.
Security scanned
Checked by our 8-point scan before listing
Fresh listing
Recently published to Agensi
Money-back guarantee
Doesn't do what its page promises? You get your money back.
Trust & safety
Security scan
Security scan passed 3 days ago
- Software
python 3.1+langgraph and langchain-core (pip install)
- Optional
- OllamaCrewAI
On Windows, use a virtual environment and run the test commands from PowerShell.
Permissions
File Scopes
Terminal is used to run the offline tests and the evaluation harness. Files are read and written only inside the skill's own project folder.