The architecture theory behind the production SDLC system
Production Track 01. A written guide extracted from the closed-source Agentic SDLC pipeline, documenting design decisions, trade-offs, and the patterns that survived contact with real users.
Agentic Patterns is the documentation half of the open-source release. It's the answer to “why did you build the production agent system the way you did?” without leaking anything proprietary. If you're evaluating me for a Forward Deployed role and want to read how I think about agent architecture before reading any code, start here.
What the guide covers
The architectural concepts are illustrated with two interactive reference architectures detailing core agent patterns (featuring live CSS keyframe execution tracing):
Five Patterns : One Architecture
Each node is a documented design decision. The connections show how they reinforce each other.
The Five Patterns : Annotated Reference
Each pattern, the problem it solves, and the decision that survived production.
Loop vs DAG execution. When an agent system should be a single recursive loop versus a topologically-ordered DAG. The honest answer: most production systems converge to a DAG with one or two bounded loop sub-graphs.
Multi-provider failover routing. Routing LLM traffic across Anthropic, OpenAI, and a local Ollama fallback. Circuit-breaker placement, retry budgets, and which signals justify re-routing versus retry.
Data drift avoidance. How the planner-implementer-reviewer agent triplet avoids the “each step amplifies the previous step's noise” failure mode. Hard structured outputs at every boundary.
Prompt-injection defences. Treating external content (repo files, customer specs, user messages) as untrusted input by default. Tagging, isolation, and the cases where defence-in-depth is actually required.
Cost and quota architecture. Where to put the SHA-256 idempotency cache, when to enforce per-user quota, and how to bound worst-case spend under adversarial input.
Why this exists as a separate track
The production system this is extracted from has paying beta users and proprietary prompt content. I can't open-source the pipeline itself. But the architectural decisions are the part that's worth reading, and they're not proprietary : they're just specific. So I wrote them up in their own track with clean reference architectures and no API keys.
Who this is for
Recruiters / hiring engineers evaluating whether I can reason about real systems beyond the code surface. This track is the fastest way to get a read on that.
Engineers building their own agent systems who want to see one concrete set of decisions written up : not as best practices, but as “here's what one person learned shipping a production system, here's what they'd do differently.”
FDE candidates entering the field : the audience this site was partly written for. The patterns are deliberately accessible.
Honest framing
This is documentation, not code. The companion track : Agentic Systems : is the runnable scaffold set. Read both. The theory without the code is incomplete; the code without the theory is just a folder of scripts.