AI workflow engineering

AI makes execution fast. I built the operating layer that makes it trustworthy.

CalvinOS turns product evidence into reviewable work without collapsing source truth, human judgment, external authority, and verification into one opaque “agent completed” state.

Inspect the real run

One bounded requestProduct Content Ops

01Route sources
02Build artifact
03Gate authority
04Verify truth

Only accepted facts become durable context.

Role: Designer and builder
Scope: Orchestration, source routing, review, verification
Status: Actively used internal system
Proof: One sanitized Product Content Ops run

Technical surface

Agent routing
MCP integrations
Approval gates
Validators and audit trail

The problem

When output gets cheap, coordination debt gets expensive.

Agentic coding made it possible to research, draft, implement, and verify work much faster. It also made a dangerous failure mode easier: an agent could produce a plausible result while using stale context, crossing an authority boundary, or promoting an unverified claim.

The hard problem was no longer making an agent perform a task. It was keeping accelerated actions aligned with the right source, exact scope, review state, and owner.

I built CalvinOS as a human-readable operating layer. Product repositories still own implementation truth. Live systems still own dynamic state. CalvinOS routes evidence, creates review artifacts, records decisions, and promotes only accepted facts into durable context.

Route truthUse the system that actually owns each fact.

Show the workMake exact output inspectable before it advances.

Gate authoritySeparate approval, execution, and cleanup.

Keep uncertaintyPreserve partial and blocked states as useful truth.

The real run

One Product Content Ops run, opened state by state.

The trace follows a real kardfolio campaign workflow. Private details are removed; the sequence, review outcome, authority boundary, and evidence gaps are preserved.

Run trace / 01Bounded

Create one kardfolio product campaign from current, claim-safe evidence.

The workflow started with one product and one campaign concept. It did not fan out across every product or personal account.

Source / owner: Calvin / Product Content Ops
Observed: Jul 27, 2026
Allowed next: Read only the permitted product sources
Completion evidence: One product and one campaign concept recorded

Run trace / 02Read only

Route product truth, public release evidence, and approved synthetic UI into separate evidence lanes.

Bounded scouts could inspect evidence in parallel. Product facts still belonged to the product repository and public surfaces—not to an agent’s memory.

Source / owner: Product repository + approved product context
Observed: Jul 27, 2026
Allowed next: Freeze the source and claim manifest
Completion evidence: Source inventory preserved with explicit gaps

Run trace / 03Reviewable

Assemble exact X, Threads, and Instagram variants with an asset and claim manifest.

The output was a human-readable review package, not an invisible agent handoff. Every external variant could be inspected as an exact revision.

Source / owner: Product Content Ops
Observed: Jul 27, 2026
Allowed next: Submit the exact revision for independent review
Completion evidence: Three platform variants and review assets assembled

Run trace / 04Revised

Run independent voice, claim, scheduling, and visual checks before approval.

The first visual review requested a crop change. The package advanced only after the corrected revision passed the same review surface.

Source / owner: Voice, scheduling, claim, and visual reviewers
Observed: Jul 27, 2026
Allowed next: Return only the changed output to review
Completion evidence: First review requested revision; verification passed

Run trace / 05Human gate

Approve the exact content without silently granting scheduling or publishing authority.

Content approval, native scheduling, and publication evidence remained distinct states. Calvin later confirmed scheduling; CalvinOS reconciled that fact rather than performing the external action.

Source / owner: Calvin
Observed: Jul 27, 2026
Allowed next: Grant separate, scoped scheduling authority
Completion evidence: Content approval recorded; external execution remained unavailable

Blocked transitionTransition approved content to scheduled

Blocked: content approval did not include scheduling or publication authority.

Run trace / 06Partial

Preflight attributed links and preserve missing external object IDs as an explicit evidence gap.

The final links returned HTTP 200. Platform object IDs were not captured, so the record remained partial instead of being promoted to a stronger state by inference.

Source / owner: Product Content Ops run record
Observed: Jul 27, 2026
Allowed next: Capture native platform object IDs before stronger verification
Completion evidence: Campaign links passed preflight; native object IDs remain missing

Select a stage to inspect its evidence and authority state.

Content approval, native scheduling, and publication evidence remained separate states. CalvinOS reconciled the confirmed external action; it did not perform it.

Failure and learning

The useful test was what happened when a boundary failed.

Detected

A local capture inherited an external analytics configuration.

Synthetic test events left the intended local boundary.

Recorded

The workflow preserved the affected scope and incident state.

It did not hide the failure behind a successful content artifact.

Blocked

Cleanup required separate authority.

The system refused to self-delete external data without approval.

The strengthened control is simple: local synthetic validation must prove external analytics are disabled before capture begins. Cleanup remains a distinct, auditable action.

Transfer to product work

The architecture is personal. The operating ideas are organizational.

CHAT is a separate native Personal Superapp—not a direct CalvinOS client. It shows how the same discipline becomes product architecture: explicit ownership, typed domains, server-enforced access, evidence references, and replay-safe mutations.

System of recordSupabase owns persisted state.
Explicit ownershipPersonal and Household access is modeled.
Server enforcementRLS uses authenticated identity.
Safe mutationEvidence and idempotency keep changes explainable.

On a product team, the labels would change—tickets, pull requests, design reviews, release approvals—but the useful structure transfers: source owners, reviewable artifacts, narrow authority, verified side effects, and learning that does not overwrite live truth.

CHAT Inventory screen in Traditional Chinese showing a demo Pencil item with its status and location — Real native proof / Inventory
**Source identity** → **Typed record** → **Personal ownership** → **RLS** → **Native UI**

Real native proof / Inventory
**Source identity** → **Typed record** → **Personal ownership** → **RLS** → **Native UI**

Next caseInspect kardfolio