For developers building on CrewAI

Production Agent Starter Kit

A CrewAI scaffold that's ready to deploy, not ready to demo.

Tutorials get you a crew that runs once on your laptop. This gets you one that runs unattended — with logs you can search, retries that survive a flaky API, and a container that deploys.

Join the Waitlist See What's Inside

The gap between a demo and production

Agent projects rarely fail because the model wasn't smart enough. They fail on cost, on debuggability, and on brittleness in the plumbing around the model.

The bill nobody planned for

An agent makes one call per step and decides how many steps to take. A scheduled crew looping on an impossible task over a weekend is the classic incident.

No way to debug it

A crew produces a bad answer at 3am. "Task failed" tells you nothing — which agent, which tool, what arguments, what came back?

One flaky call kills the run

A tool 502s and the whole run dies, throwing away every task that already succeeded and everything they cost.

What's actually in the repo

Six things, each in a file you can read in a couple of minutes.

Config-driven agents

Agents and tasks live in config/agents.yaml and config/tasks.yaml. Changing a crew never means editing Python — the thing almost every tutorial hardcodes.

config/*.yaml

Structured logging

One JSON object per line for every agent action, tool call, and task outcome, taken from CrewAI's own event bus. Point it at Datadog or Loki and you're done.

src/observability.py

Retries that know what to retry

Exponential backoff with jitter on transient failures. A 404 fails immediately instead of costing you three pointless retries.

src/tools/base.py

Failures that don't kill the run

A dead tool degrades one step rather than throwing away a run that already cost money. The agent is told what it can't rely on, and adapts.

src/tools/base.py

Fail-fast config

A missing API key crashes in about a second with a message naming the variable — not ninety seconds into a paid run, after two tasks completed.

src/settings.py

Tests that actually run

26 tests, no API key, no network, a quarter of a second. The model is stubbed through CrewAI's documented extension point, so tests hit the real code path.

tests/

The code is the product

So here's some of it.

Define an agent — config/agents.yaml

researcher:
  role: Research Analyst
  goal: >
    Extract accurate, verifiable facts from
    source material, and be explicit about
    what the source does not say.
  backstory: >
    You have spent years doing desk research
    where being wrong is expensive. You quote
    figures exactly as written.
  tools:
    - fetch_webpage
  max_iter: 8

No Python edited. No rebuild. No restart.

What it logs — LOG_FORMAT=json

{
  "timestamp": "2026-07-31T12:41:42+0100",
  "level": "INFO",
  "message": "fetch_webpage succeeded",
  "event": "tool.finished",
  "agent": "Research Analyst",
  "tool": "fetch_webpage",
  "outcome": "success",
  "duration_ms": 812.4
}

Every tool call, task, and retry. Searchable by field.

Two tiers

Both tiers include the entire kit. The only difference is how many developers the licence covers.

In every tier

Full source: config-driven crew, tools, logging, retries

Production notes — what breaks in production and what to do about it

26-test offline suite

Dockerfile + docker-compose

Fly.io deploy script and walkthrough

README and quickstart guide

Video walkthrough

Pricing is being finalised. Waitlist members get launch pricing and first access.

Solo

One developer. Unlimited projects.

Everything in the kit — nothing held back
Licensed to you as an individual
Unlimited personal, commercial, and client projects
Deploy to production and ship it inside software you sell
Lifetime updates

Team

Most useful

Up to 10 developers in one organisation.

Everything in Solo, for your whole team
Up to 10 developers at any one time
Use across unlimited client engagements
Hand an End Product to a client with no licence of their own
Lifetime updates

Get it when it ships

One email when it's ready to buy, with launch pricing. Nothing else in between.

No spam, unsubscribe anytime. Irish company, GDPR-compliant.

Questions

Is this the same as Devlar AI Workforce?

No. AI Workforce is our autonomous agent platform for businesses. This is a developer product — a repo you clone and build on. Separate thing, separate purchase.

Do I need to know CrewAI already?

Some familiarity helps, but the quickstart assumes you're new to it. The code is heavily commented where it teaches something, because the code is the product.

Which LLM providers work?

OpenAI, Anthropic, Groq, Gemini, and Mistral, switched with one environment variable. It routes through LiteLLM, so adding another provider is a one-line change.

What's the difference between Solo and Team?

Only the number of developers the licence covers — one, or up to ten in a single organisation. Both tiers get the identical repo, docs, and video. Nothing is held back from Solo.

Can I use it in client work?

Yes — unlimited personal and commercial projects, including software you sell, and unlimited client engagements. You can hand a finished product to a client without them buying their own licence. The only restriction is reselling the kit itself as a competing boilerplate, template, or course.

What are the requirements?

Python 3.10 to 3.13, and an API key for one LLM provider. Docker is optional and only needed for the container workflow.