Plan before code
We write down goals, risks, and user value before agent flows implement.
Careers and culture
How we work with agentic workflows
Caire builds an AI-first healthcare platform, but with senior product discipline: small verified changes, clear owners, human review, and respect for real home care.
Our way of building
The same principle used in Caire Core applies internally: a continuous learning loop, clear decisions, traceability, and human responsibility before automation.
Verification at every step
AI can produce a lot, but code, design, content, and data flows are reviewed as product work.
Care first
Technology is judged against real home-care work: planners, caregivers, clients, and family.
Small vertical releases
We prefer useful improvements that can be tested, measured, and improved over large opaque leaps.
Delivery mandate
Humans define what. Agents execute how.
Product owners write a PRD with change, success metric, and non-goals. Specialized agents then handle Spec, tests, implementation, review, verification, dossier, and PR flow while humans stay accountable for the product decision.
North star
features per second per dollar
We measure the elapsed time from PRD intake to merged PR, model spend per shipped feature, and token efficiency. The point is not replacing product judgment; it is removing avoidable handoffs from work that has already been specified.
Workflow
Eight stages, PRD to PR
Each stage produces a checked artifact. The next stage refuses to start without it, making the pipeline a one-way contract from product intent to merged change.
| # | Stage | Output | Owner |
|---|---|---|---|
| 0 | Intake normalize | Canonical PRD path and feature slug | Orchestrator |
| 1 | PRD | Change, success metric, non-goals | Human |
| 2 | Spec | Gherkin acceptance scenarios | Architect agent |
| 3 | Tests | Failing Vitest and Playwright tests | Test-writer |
| 4 | Implementation | Code in an isolated worktree | Editor |
| 5 | Review | P1/P2/P3 findings or clean report | Reviewer agents |
| 6 | Verification | Trace, screenshot, summary dossier | Verifier |
| 7 | PR | Pushed branch and external review | GitHub flow |
Migrated platform source
The detailed operating model behind the culture page
The old platform documentation had separate subpages for each agentic workflow concept. Those pages are now folded into this Brand 2.0 careers page so candidates and partners can understand the system without leaving the main website.
Vision and mandate
The job is to shorten the path from product intent to verified change without removing accountability. Read this first: humans own priority, quality, and decisions.
SourceThe job descriptionAbout CaireThe role
Agent roles and model routing
The workflow splits responsibility across Architect, Editor, Reviewer, Verifier, and Orchestrator. The role chooses the model by task, cost, risk, and context.
The five rolesCost rationaleRouting matrixPer-role responsibilities
Darwin as orchestrator
Darwin is a thin in-house Node loop that holds intake, queue, status, and evidence together. It should stay simple, observable, and replaceable in parts.
End-to-end flowA single orchestratorOrchestrator runtimeRunner resilience
Darwin component map
The component map separates what is used by the PRD-to-PR pipeline, what is parked, and what remains in the prioritized build queue.
Used by PRD-to-PR pipelineParkedOutstandingTelegram routing
Model and vendor agnosticism
The adapter shape lets the workflow select a model without locking the product process to one vendor or one interface.
The ruleAdapter shapeWhat this rules outWhat this allows
Spec as contract
Spec means testable acceptance scenarios, not loose wishes. Gherkin and a done contract gate reduce spec drift between PRD, tests, and implementation.
GherkinDone contract gateSingle-source-of-truthSpec drift
Reviewer feedback loop
The review loop polls external reviewers, classifies severity, and re-enters the editor with an idempotent action list.
The pollSeverity → actionRe-entering the editorShort-circuit
Scale or kill
Automation should scale only when evidence, cost, and quality hold. Otherwise it must be easy to stop or roll back.
ScaleKillAuto-promoteAuto-rollback
Throughput and business signals
The features per second per token metric is read together with cost, quality, and business signals to see whether the workflow actually improves delivery.
Features per second per tokenWhere it's loggedWho reads itBusiness signals
Verification and evidence
Every delivery should leave a trace: trace.zip, screenshot.png, console.log, test excerpts, and a short explanation of what the artifact proves.
docs/dossiers/<feature>/trace.zipscreenshot.pngconsole.log
Caire Core
The product we build and the way we build it share a philosophy: trace what changed, learn from outcomes, and keep humans accountable for decisions that matter.