feat: add kalibr-resilient-crew — production reliability layer for CrewAI agents

crews/kalibr-resilient-crew/.env.example

@@ -0,0 +1,13 @@
+# OpenAI — primary model
+OPENAI_API_KEY=your_openai_api_key_here
+
+# Anthropic — fallback model (Kalibr routes here if GPT-4o degrades)
+ANTHROPIC_API_KEY=your_anthropic_api_key_here
+
+# Serper — web search tool
+SERPER_API_KEY=your_serper_api_key_here
+
+# Kalibr — execution path routing
+# Get credentials: https://dashboard.kalibr.systems or run `kalibr auth`
+KALIBR_API_KEY=your_kalibr_api_key_here
+KALIBR_TENANT_ID=your_kalibr_tenant_id_here

crews/kalibr-resilient-crew/README.md

@@ -0,0 +1,50 @@
+# Kalibr-Resilient Job Posting Crew
+
+A production-grade CrewAI crew with automatic LLM routing and failure recovery via [Kalibr](https://kalibr.systems).
+
+This example extends the standard job-posting crew with Kalibr's execution path router. The crew researches company culture, identifies role requirements, drafts a job posting, and reviews it — all while Kalibr monitors outcome signals and routes each LLM call to the model most likely to succeed.
+
+## Why Kalibr
+
+Most CrewAI examples are hardcoded to GPT-4o. That's fine for demos. In production:
+
+- Provider degradation events (rate limits, latency spikes, partial outages) happen multiple times per week
+- A single-model crew fails when that model degrades
+- Kalibr routes to your backup model automatically, learned from live outcome data
+
+**Benchmark results during a simulated GPT-4o degradation event:**
+| Setup | Task Success Rate |
+|---|---|
+| Hardcoded GPT-4o | 16–36% |
+| Kalibr-routed (GPT-4o + Claude Sonnet) | 88–100% |
+
+## Running the Script
+
+**Configure Environment**: Copy `.env.example` to `.env` and fill in:
+- `OPENAI_API_KEY` — [OpenAI API key](https://platform.openai.com/api-keys)
+- `ANTHROPIC_API_KEY` — [Anthropic API key](https://console.anthropic.com/) (fallback model)
+- `SERPER_API_KEY` — [Serper](https://serper.dev) for web search
+- `KALIBR_API_KEY` and `KALIBR_TENANT_ID` — [Kalibr dashboard](https://dashboard.kalibr.systems)
+
+**Install Dependencies**: `uv sync`
+
+**Execute**: `uv run kalibr_resilient_crew`
+
+## How Kalibr works here
+
+`import kalibr` at the top of `main.py` (before any OpenAI/Anthropic import) monkey-patches the SDK clients. From that point on, every LLM call the crew makes is:
+1. Traced with task goal context
+2. Routed to the currently-succeeding model
+3. Scored on outcome (did the agent produce valid output?)
+4. Used to improve future routing decisions
+
+No changes to agent definitions, task configs, or crew logic.
+
+## Get Kalibr credentials
+
+```bash
+pip install kalibr
+kalibr auth  # device-code auth — opens browser, one click
+```
+
+Or sign up at [dashboard.kalibr.systems](https://dashboard.kalibr.systems).

crews/kalibr-resilient-crew/pyproject.toml

@@ -0,0 +1,16 @@
+[project]
+name = "kalibr_resilient_crew"
+version = "0.1.0"
+description = "Production-resilient job posting crew with Kalibr execution path routing"
+authors = [{name = "Devon Kelley", email = "devon@kalibr.systems"}]
+readme = "README.md"
+requires-python = ">=3.12,<=3.13"
+dependencies = [
+    "crewai[tools]>=0.152.0",
+    "python-dotenv>=1.0.1",
+    "kalibr>=0.1.0",
+]
+
+[project.scripts]
+kalibr_resilient_crew = "kalibr_resilient_crew.main:run"
+train = "kalibr_resilient_crew.main:train"

crews/kalibr-resilient-crew/src/kalibr_resilient_crew/__init__.py

@@ -0,0 +1,2 @@
+# Kalibr MUST be imported before any OpenAI/Anthropic/LLM imports.
+# This file is intentionally minimal — see main.py for entry point.

crews/kalibr-resilient-crew/src/kalibr_resilient_crew/config/agents.yaml

@@ -0,0 +1,29 @@
+research_agent:
+  role: >
+    Research Analyst
+  goal: >
+    Analyze the company website and provided description to extract
+    insights on culture, values, and specific needs.
+  backstory: >
+    Expert in analyzing company cultures and identifying key values
+    and needs from various sources, including websites and brief descriptions.
+
+writer_agent:
+  role: >
+    Job Description Writer
+  goal: >
+    Use insights from the Research Analyst to create a detailed,
+    engaging, and enticing job posting.
+  backstory: >
+    Skilled in crafting compelling job descriptions that resonate
+    with the company's values and attract the right candidates.
+
+review_agent:
+  role: >
+    Review and Editing Specialist
+  goal: >
+    Review the job posting for clarity, engagement, grammatical accuracy,
+    and alignment with company values and refine it to ensure perfection.
+  backstory: >
+    A meticulous editor with an eye for detail, ensuring every piece of content
+    is clear, engaging, and grammatically perfect.

crews/kalibr-resilient-crew/src/kalibr_resilient_crew/config/tasks.yaml

@@ -0,0 +1,36 @@
+research_company_culture_task:
+  description: >
+    Analyze the provided company website {company_domain} and the following
+    company description: {company_description}.
+    Focus on identifying the company's culture, values, and mission.
+  expected_output: >
+    A comprehensive summary of the company's culture and values.
+
+research_role_requirements_task:
+  description: >
+    Based on the hiring needs: {hiring_needs}, identify the skills,
+    experience, and qualities required for the role.
+  expected_output: >
+    A structured JSON with skills, experience, and qualities lists.
+
+draft_job_posting_task:
+  description: >
+    Draft a job posting for {hiring_needs} at a company with the following
+    description: {company_description}.
+    Highlight the specific benefits: {specific_benefits}.
+  expected_output: >
+    A detailed, engaging job posting ready for publication.
+
+review_and_edit_job_posting_task:
+  description: >
+    Review the job posting for clarity, engagement, grammatical accuracy,
+    and alignment with company values.
+  expected_output: >
+    A polished, final job posting ready for immediate publication.
+
+industry_analysis_task:
+  description: >
+    Conduct an industry analysis for the role: {hiring_needs}.
+    Identify salary ranges, required certifications, and market trends.
+  expected_output: >
+    A comprehensive industry analysis with salary benchmarks and market insights.

crews/kalibr-resilient-crew/src/kalibr_resilient_crew/crew.py

@@ -0,0 +1,69 @@
+from typing import List
+
+from crewai import Agent, Crew, Process, Task
+from crewai.project import CrewBase, agent, crew, task
+from crewai_tools import SerperDevTool, WebsiteSearchTool, FileReadTool
+from pydantic import BaseModel, Field
+from kalibr.integrations.crewai import KalibrCrewAIInstrumentor
+
+import os
+
+KalibrCrewAIInstrumentor(
+    tenant_id=os.environ.get("KALIBR_TENANT_ID", ""),
+    goal="job_posting_production",
+).instrument()
+
+web_search_tool = WebsiteSearchTool()
+serper_dev_tool = SerperDevTool()
+file_read_tool = FileReadTool(
+    file_path='job_description_example.md',
+    description='A tool to read the job description example file.',
+)
+
+
+class ResearchRoleRequirements(BaseModel):
+    skills: List[str] = Field(..., description="List of recommended skills for the ideal candidate.")
+    experience: List[str] = Field(..., description="List of recommended experience for the ideal candidate.")
+    qualities: List[str] = Field(..., description="List of recommended qualities for the ideal candidate.")
+
+
+@CrewBase
+class KalibrResilientJobPostingCrew:
+    agents_config = 'config/agents.yaml'
+    tasks_config = 'config/tasks.yaml'
+
+    @agent
+    def research_agent(self) -> Agent:
+        return Agent(config=self.agents_config['research_agent'], tools=[web_search_tool, serper_dev_tool], verbose=True)
+
+    @agent
+    def writer_agent(self) -> Agent:
+        return Agent(config=self.agents_config['writer_agent'], tools=[web_search_tool, serper_dev_tool, file_read_tool], verbose=True)
+
+    @agent
+    def review_agent(self) -> Agent:
+        return Agent(config=self.agents_config['review_agent'], tools=[web_search_tool, serper_dev_tool, file_read_tool], verbose=True)
+
+    @task
+    def research_company_culture_task(self) -> Task:
+        return Task(config=self.tasks_config['research_company_culture_task'], agent=self.research_agent())
+
+    @task
+    def research_role_requirements_task(self) -> Task:
+        return Task(config=self.tasks_config['research_role_requirements_task'], agent=self.research_agent(), output_json=ResearchRoleRequirements)
+
+    @task
+    def draft_job_posting_task(self) -> Task:
+        return Task(config=self.tasks_config['draft_job_posting_task'], agent=self.writer_agent())
+
+    @task
+    def review_and_edit_job_posting_task(self) -> Task:
+        return Task(config=self.tasks_config['review_and_edit_job_posting_task'], agent=self.review_agent())
+
+    @task
+    def industry_analysis_task(self) -> Task:
+        return Task(config=self.tasks_config['industry_analysis_task'], agent=self.research_agent())
+
+    @crew
+    def crew(self) -> Crew:
+        return Crew(agents=self.agents, tasks=self.tasks, process=Process.sequential, verbose=True)

crews/kalibr-resilient-crew/src/kalibr_resilient_crew/main.py

@@ -0,0 +1,51 @@
+# IMPORTANT: kalibr must be the first import.
+# It monkey-patches OpenAI and Anthropic SDK clients so that every
+# LLM call in this process is automatically traced, routed, and scored.
+import kalibr  # noqa: F401 — must be first
+
+import sys
+from dotenv import load_dotenv
+from kalibr_resilient_crew.crew import KalibrResilientJobPostingCrew
+
+load_dotenv()
+
+
+def run():
+    """
+    Run the Kalibr-resilient job posting crew.
+
+    Kalibr automatically routes each LLM call to the optimal model based on
+    live outcome signals. If GPT-4o degrades, calls route to Claude Sonnet.
+    No code changes needed — routing adapts in real time.
+    """
+    inputs = {
+        'company_domain': 'careers.wbd.com',
+        'company_description': (
+            "Warner Bros. Discovery is a premier global media and entertainment company, "
+            "offering audiences the world's most differentiated and complete portfolio of "
+            "content, brands and franchises across television, film, sports, news, "
+            "streaming and gaming."
+        ),
+        'hiring_needs': 'Production Assistant, for a TV production set in Los Angeles in June 2025',
+        'specific_benefits': 'Weekly Pay, Employee Meals, healthcare',
+    }
+    KalibrResilientJobPostingCrew().crew().kickoff(inputs=inputs)
+
+
+def train():
+    """Train the crew for a given number of iterations."""
+    inputs = {
+        'company_domain': 'careers.wbd.com',
+        'company_description': (
+            "Warner Bros. Discovery is a premier global media and entertainment company."
+        ),
+        'hiring_needs': 'Production Assistant, for a TV production set in Los Angeles',
+        'specific_benefits': 'Weekly Pay, Employee Meals, healthcare',
+    }
+    try:
+        KalibrResilientJobPostingCrew().crew().train(
+            n_iterations=int(sys.argv[1]),
+            inputs=inputs,
+        )
+    except Exception as e:
+        raise Exception(f"An error occurred while training the crew: {e}")