Didur Labs software delivery operating system

AI agents with a real software process.

This page explains how Ryan Didur's setup turns ideas into planned, reviewed, and deployed applications using Hermes, Archer, specialist agents, GitHub, Rancher, and Kubernetes — without vibe coding or using the Hermes server as an app host.

Architect first

Archer designs systems and writes sprint packs before builders code.

Project docs

Each project has its own source-of-truth planning folder and repo docs.

Evidence gates

Implementation, tests, review, deployment, and health checks are required.

Rancher runtime

Production apps run on Rancher/Kubernetes, not on Hermes.

The core rule

Hermes is the agent home. It coordinates, remembers, schedules, and routes work. It is not the place where production applications live, and it is not the long-term development workstation for app source.

Allowed: Hermes orchestration, agent profiles, Kanban, cron, skills, memory, automation.
Source of truth: GitHub repos, project docs, sprint packs, reviewed commits.
Runtime: Rancher/Kubernetes with ingress, TLS, health checks, and rollback evidence.
Not allowed: app source, dev servers, previews, or production apps living on Hermes.

Why this exists

AI can produce code quickly, but speed without structure creates context drift, undocumented decisions, weak testing, and fragile deployments. The system uses AI where it helps — planning, coding, review, operations — while preserving normal software engineering controls.

No vibe codingArchitecture packsTDD when practicalReview gatesRancher deploysPublic health evidence

End-to-end workflow

Every application follows a pipeline. Agents can run autonomously, but the handoffs and evidence stay durable.

1

Intake

Ideas, inbox specs, meetings, voice notes, and research land in Dropbox/Hermes/SoftwareProcess/03-intake or a Kanban row.

2

Architecture

Archer inventories the request, updates the project folder, defines architecture, and writes sprint requirements, blueprint, acceptance criteria, and builder handoff prompt.

3

Implementation

Builders such as Cody, Mason, Rowan, and Elena work from repo/worktree source, not chat history. They commit code and test evidence.

4

Review

Quinn reviews against acceptance criteria and evidence. Failed reviews produce focused fix tasks, not vague retries.

5

Deployment

Atlas deploys production workloads to Rancher/Kubernetes with namespace, workload, ingress, TLS, health, and rollback evidence.

6

Operations

Piper keeps the board high-signal. Sable, Ada, Nora, Linus, Victor, Hazel, and other specialists own security, identity, networking, Linux, VMware, and smart-home lanes.

Per-project documentation

Every active project is organized the same way under Dropbox/Hermes/SoftwareProcess/02-projects and mirrored into repo docs where appropriate.

Required planning files

README.mdAGENTS.mdplanning/state.mdplanning/decisions.mdplanning/domain.mdplanning/risks.mdplanning/questions.mdplanning/file-inventory.md

Required engineering files

docs/architecture.mddocs/data-model.mddocs/api.mddocs/permissions.mddocs/validation.mdsprint requirementssprint blueprintacceptance criteriahandoff prompt

Agent roles

Agent/laneOwnsProducesEvidence before done
ArcherSoftware architecture and portfolio designArchitecture docs, sprint packs, repo/Rancher decisionsUpdated project docs and builder handoff prompt
BuildersImplementationCode, tests, commits, PRsRepo branch/commit, test output, docs updates
QuinnQA and reviewPass/fail review, defects, regression notesAcceptance-criteria review and reproduction evidence
AtlasRancher/Kubernetes deploymentManifests, ingress, TLS, CI/CD, health checksNamespace/workload/ingress/TLS/URL health evidence
PiperOrchestration and board hygieneMinimal task graph, blocker cleanup, status relayReal task IDs, clear dependencies, stale rows closed
Sable/Ada/Nora/Linus/Victor/HazelSecurity, identity, DNS/network, Linux, VMware, Home AssistantSpecialist changes and domain evidenceTool-backed validation, reports, and safe handoffs

Training program

New software agents are trained against the same operating standard before touching builds. The training covers Ryan's non-negotiables, architect-builder workflow, per-project documentation, build discipline, review, deployment, and board hygiene.

06-agent-training/curriculum.mdonboarding checklistrole matrixgraduation prompt

Current project areas

AITraderITOperationsHQSaaSMonHermes M365 OAuth connectorHome Assistant OpenAI/Bifrost integrationValidation apps

Each project gets its own folder, docs, sprint packs, repo decision, and deployment path.