Stochastic generation can’t go in your SDLC. Drift, hallucination, and an audit trail full of prompts aren’t answers your security review will accept.
nmux is a deterministic, locally-run pipeline of specialized agents, each gated by an enforceable contract. The outputs are reproducible, auditable, and SHA-256 verifiable. This page is the architectural case for adopting it.
Five phases, five contracts. Or ten, or twenty. You decide. Your SDLC requirements dictate how things get built, every time all the time. The output of one is the verified input of the next. Nothing advances without passing its gate. Hyper engineering, by construction.
Got some pesky compliance gates that must be passed before code can be merged? Model them as a phase. Want to enforce a test-driven development workflow? Model it as a phase. Want to make sure every spec is reviewed by a senior engineer before implementation begins? Model it as a phase. Your process, your rules, encoded in the architecture of your pipeline.
Single-shot review by a generalist model is the wrong shape for production code. nmux ships a default swarm — each agent owns one criterion, each is content-addressable, each emits a typed finding the gate can act on. Swap any for one of yours. Or define your own.
Every nmux run produces a content-addressable artifact tree. ADRs, contracts, implementation, tests, agent findings, audit trail, attestation.
1 {
2 "phase": "review",
3 "agents": 8,
4 "blocking": 0,
5 "advisory": [
6 {
7 "agent": "performance",
8 "severity": "advisory",
9 "location": "src/order.ts:142",
10 "criterion": "hot-path budget",
11 "message": "O(n²) on large carts; consider memoization",
12 "rule": "perf.complexity.advisory"
13 },
14 { "agent": "style", "severity": "advisory", "count": 2 }
15 ],
16 "verdict": "pass-with-advisories"
17 }
Same spec → same pipeline, same gates. Re-run six months from now and the process is identical. The non-determinism of LLMs is contained inside agent boundaries, not exposed to the SDLC.
Distributed as a single static binary. No telemetry, no calls home, no third-party SaaS in your supply chain. Code, prompts, and outputs all stay in your network — no asterisk.
ADRs, contracts, and house policies aren't aspirational README content — they're contracts the gate checks. The pipeline halts on violation.
Every routing decision, every agent invocation, every finding — logged, hashed, attributed. You can answer 'why was this code generated' six months later. With evidence.
We’ve sat through enough architecture reviews to know which questions come up. Here they are, with the short version of each answer.
Single-model review drifts — the one that's strong at types is weak at perf; the one that catches injection misses ADR fit. Specialists are criterion-bound, content-addressable, and produce typed findings the gate can act on. Swap any. Define your own.
Agents run in parallel by design — wall-clock is bounded by the slowest, not the sum. Inference cost scales with the compute you provision, not seat count. The model in the business case estimates the compute load for a typical 14-engineer team.
They don't. Each phase is contract-bound — halt-on-violation, not silently-degrade. Findings are typed and attributed to a specific agent. State persists across sessions, so failures resume from the last checkpoint, not from scratch.
It IS your CI/CD action. First-class GitHub Action, shell-out for CircleCI/Buildkite, IDE plugins for VS Code/Cursor/JetBrains. Same artifact bundle regardless of where you run it. See §06 above for the integration matrix.
30 minutes with one of our engineers. We’ll run a representative spec through the full pipeline, live, and hand you the artifact bundle that lands in your repo. Bring your own spec if you want.