Agentic Engineering Labs

Twelve runnable, test-driven projects for the Organized AI Agentic Engineering field guide.

Every project includes:

• README.md — objective, concepts, worked example, build steps, and extension.
• starter.py — a deliberately incomplete implementation with marked TODOs.
• solution.py — a small reference implementation.
• test_solution.py — executable checkpoints for the reference and your work.

The core track uses Python 3.10+ and the standard library only. It makes no network calls, needs no API keys, and uses synthetic data. Advanced GPU and provider integrations are documented as optional extensions.

Quick start

git clone https://github.com/Organized-AI/agentic-engineering-labs.git
cd agentic-engineering-labs
python3 scripts/test_all.py

The current reference suite contains 70 tests across 23 projects. GitHub Actions runs the same structure and test commands on every push and pull request.

Guide website source

The complete source for guide.organizedai.vip/agentic-eng is versioned in guide-site/. It includes chapter Markdown, static-site source, build and verification scripts, generated deployment assets, research audit records, local font licenses, and the Cloudflare Worker config.

cd guide-site
npm ci
npm run build
npm run check

node_modules/ and .wrangler/ are intentionally excluded because they are reproducible dependency caches and machine-local Cloudflare state.

Run one project:

cd projects/03-jobs-and-events
python3 -m unittest -v test_solution.py
python3 solution.py

To work from the starter, copy it outside the repository's reference path or edit it directly, then change test_solution.py to import starter instead of solution. A good rhythm is red → green → refactor: make one test fail for the right reason, implement the smallest fix, then clean up without changing the result.

The projects

#	Topic	Project	Checkpoint
01	Experimentation	Paired experiment ledger	Compare the same cases and explain regressions
02	Engineering foundations	Tenant-safe event service	Validate contracts and reject stale writes
03	Jobs & events	Leased job runner	Deduplicate, recover, and fence stale workers
04	LLM gateways	Policy-aware model router	Preserve policy and budget during fallback
05	Agent design	Bounded tool agent	Prevent untrusted text from expanding authority
06	Evaluations	Outcome evaluation harness	Grade state, repeats, and release gates
07	Inference infrastructure	Memory and hosting planner	Estimate weights/KV cache and compare ownership
08	Load testing	Queueing workload simulator	Find the goodput saturation point
09	Data retention	Synthetic canary audit	Find persistence across success and failure paths
10	Kernels & performance	Vector-kernel benchmark	Prove correctness before claiming speedup
11	Ontologies & semantics	Source-backed domain graph	Resolve canonical facts with provenance
12	Cost of cognition	Accepted-outcome ledger	Include retries, review, and utilization
13	Agentic security	Injection gauntlet	Reject tainted effects and rug-pulled tools
14	Agent identity	Identity-aware tool gateway	Verify subject, scope, and audience per call
15	Human oversight operations	Approval queue state machine	Exact, live, single-use approvals with audit first
16	Context & memory systems	Typed memory store	Keep memory types, sources, and constraints separate
17	Evals, traces & logs	Three-signal debugger	Join logs, spans, and scores on one task id
18	Guardrails in the request path	Rail pipeline	Named rails, logged decisions, fail closed
19	Adaptation economics	Adaptation lever pricer	Price prompt, RAG, and fine-tune on one ledger
20	AI Act & governance	Classification memo builder	Classify the system and map duties to artifacts
21	Tailnets & agent networking	Mesh policy checker	Prove intended paths, deny everything else
22	Engineering best practices	Foundation audit	Score the classic blocks on evidence
23	The challenge	Sprint ledger	Ship an artifact at every 7/30/60/90 gate

Progress

Use PROGRESS.md as a repository checklist. The website progress dashboard tracks four milestones per chapter in browser storage and can export/import a JSON progress file. The repository and website use the same stable project slugs.

Scope

These are instructional simulations, not production templates or compliance certifications. Passing the tests proves only the named invariants in a small, controlled environment. Re-check provider behavior, security requirements, licenses, and retention agreements for a real deployment.