A Real Three-Repo Setup
A real three-repo setup where AI agents do actual work: one defines the workflows, one turns requests into plans, and one does the work itself.
- One repo defines the workflows. One turns requests into plans. One does the work.
- The opportunity is not to merge them. It is to let them teach each other.
- The best example is a gated pipeline, not a vague agent.
- Do not mistake "agentic service" for "autonomous agent." The value is in the gates.
Three Repos, Three Jobs
These three repos map cleanly onto three layers of the bigger curriculum.
1. Defines The Workflows
This layer is good for learning what a canonical source of truth looks like. It does five things:
- defines what workflows exist
- generates tool-specific versions of each one so it works the same way in every AI client
- describes what tools and context each workflow can use through machine-readable manifests
- records run artifacts and metadata for every execution
- validates that a generated copy hasn't drifted from its canonical source
Mental model:
It answers: What workflows exist? How do they show up in each AI tool? What tools are available? Where do artifacts go? What is the source of truth versus a generated copy?
A team might do well to:
- write down a first-class convention for multi-agent run artifacts
- build eval fixtures for common workflows such as research or ideation or review
- keep experimental runtimes isolated until they prove real value
What not to force: don't adopt a heavyweight graph runtime just because it's interesting. Don't wire in full orchestration config until at least one or two artifact-mediated workflows have proven the shape works.
2. Turns Requests Into Plans
This layer is good for learning spec-driven development. A request becomes a written spec before anything gets built. It also shows:
- workflow tiers and quality gates
- artifacts as guardrails so the work isn't done until the right evidence exists
- task completion gates tied to real checks rather than a status flag
- telemetry that shows where work gets stuck
- role-specific workflows for different kinds of contributors
Mental model:
It answers how a vague ask becomes:
- a spec
- a plan
- tasks
- an implementation
- tests
- a review
- a release
The important lesson:
This layer provides the rails; the AI agents provide the engine.
That's the same idea as AI systems orchestration generally. The model is not the whole system. The system is what makes the model's work repeatable.
- use this layer's command-to-artifact mapping as a template for the workflow layer's own artifact contracts
- use its telemetry schema as a starting point for tracking agent runs more broadly
- turn its real failed runs into eval fixtures
What not to force: don't try to merge this with the workflow-definition layer. It's closer to a bootstrapping and productization kit than the same thing. And don't collapse personas and workflows and runtime agents into one concept. They're related. They are not identical.
3. Does The Actual Work
This layer is good for learning what real LLM-backed services look like in production. It has:
- pipelines with hard gates between steps
- automation built from a mix of deterministic code and model calls
- tool servers that expose safe and scoped access to real systems
- production-grade concerns such as:
- retries
- correlation IDs
- redaction
- feature flags
- observability
Mental model:
It answers what an actual agentic service looks like once it has:
- APIs
- clients
- retries
- LLM calls
- context lookup
- feature flags
- logs
The strongest example of this pattern is a support-ticket triage pipeline:
incoming ticket
-> input processor
-> hypothesis generator (hard gate)
-> code-area identifier (additive context)
-> fix recommender
-> action generator
-> output formatter
-> posted back as a comment
That's a real LLM pipeline. It is not a vague "agent." It has gates, structured outputs, additive context, idempotent updates, redaction, and observability.
- use a pipeline like this as a study case for gated LLM workflows
- use a code-refactoring service as a study case for model routing and deterministic-plus-LLM automation
- use a well-scoped MCP tool server as a study case for production tool design
- build eval suites around the triage pipeline to check whether a known issue produces the expected hypothesis and the expected fix and the expected action
What not to force: don't assume every service like this needs multi-agent orchestration. Many should stay as clean linear pipelines. And don't mistake an agentic service for an autonomous agent. The best parts here are the gates and the wrappers around the model. Unconstrained autonomy is not the point.
The Three-Layer Map
A useful way to think about the three layers:
- one is the control plane
- one is the workflow scaffolding kit
- one is the runtime and service playground
The opportunity isn't to merge them into one system. It's to let them teach each other.