Engineering Agent Workflows with Persistent State and Human Review

I am building several related systems for venture execution, editorial work, and personal coordination. The shared engineering problem is maintaining a reliable connection between an instruction, its source context, the plan it produces, and the artifact a person eventually approves.

Venture Command and ContentOS implement different parts of that workflow. Aegis supplies a separate evaluation and evidence layer. Their maturity differs: local workflow implementations, synthetic tests, and verified integration paths do not establish an autonomous production operation.

Venture Command turns context into bounded work

Venture Command is a local-first workstation with a Python backend, SQLite persistence, and a buildless JavaScript dashboard. A mission begins with a goal and saved project knowledge. The system assembles relevant sources, requests a structured plan through Codex CLI, and validates the result before allowing work to proceed.

Knowledge intake accepts text, Markdown, files, and selected conversation exports. Retrieval is lexical; this implementation does not depend on a vector database. Sources are versioned so that a plan can be tied to the evidence used to produce it.

The validation layer restricts supported capabilities, task counts, and dependencies, and requires final review. If the underlying sources change, source-revision binding blocks the stale plan. This makes an otherwise subtle failure visible: a syntactically valid plan may no longer describe the current project.

Internal draft handoffs and execution limits bound the work. Durable local jobs, activity records, cancellation, backup/restore, and conservative recovery preserve operational state. An uncertain run should not be treated as a confirmed success or blindly restarted.

The dashboard exposes mission timelines and dependencies. A scoped worker API, optional MCP bridge, and portable task packets provide integration seams. Explicit Obsidian synchronization handles conflicts; structured Life Ops exports carry commitments into the coordination workflow.

The verification boundary is narrow. One synthetic live Codex planning test returned a valid two-task plan without tool calls. Broader worker orchestration was tested with mocked workers. Roles execute through one local supervisor. Repository coding, external-bot dispatch, and continuous monitoring remain incomplete; I do not present this as a deployed parallel autonomous workforce.

ContentOS binds review to the artifact

ContentOS is a project-aware editorial system. A source registry and versioned evidence packets feed deduplicated briefs and retained draft history. Agent-assisted production and independent review are coordinated within Codex sessions across editorial desks.

The important state transition is approval. Seven review checks precede exact-content hash approval. Relevant changes invalidate that approval, so a revised payload cannot silently inherit permission granted to an earlier draft.

This connects review to the actual artifact intended for delivery. A broad instruction such as “approve the post” is too ambiguous when the post, source evidence, or destination may change between review and delivery.

A separate publishing subsystem contains OAuth integration paths, macOS Keychain credential storage, scheduling, delivery records, and uncertain-send reconciliation. Cloudflare Worker infrastructure supports Meta account connection and expiring HTTPS media delivery with asset integrity checks.

The implementation and verification claims are different. Website analytics reads, Instagram test-account authentication, and synthetic media transport were verified. Those checks do not establish successful production social publishing. Unattended model-based content production remains unimplemented; Reddit and Substack use approved manual handoffs.

Analytics retains the meaning of a metric

ContentOS ingests website analytics from Cloudflare and Vercel into SQLite. Records preserve source provenance, reporting windows, and metric definitions. A macOS background collector supports recurring website analytics collection.

Keeping those fields matters because superficially similar metrics may represent different populations or time windows. The dashboard should preserve the information needed to interpret a number before comparing it with another source.

Aegis connects evaluation to deployment evidence

Aegis Conformance is a separate Python evaluation engine. Configurable claims become probes; target adapters query models or HTTP services; judges score responses; reports retain confidence intervals, baseline comparisons, and reproducibility metadata.

The deployment evidence layer binds measured results to a deployment case and human-authored authorization terms. Evidence gates distinguish a requirement measured below its threshold from a run that cannot establish whether the threshold is met. Insufficient evidence is a meaningful outcome.

Evidence bundles use SHA-256 integrity seals and optional HMAC signatures. These mechanisms support integrity checking; they do not constitute accredited certification or public-key attestation. Synthetic fixtures exercise the pipeline, and I do not claim a published frontier-model ranking from those fixtures.

Opportunity research is an application of the workflow

I also have an AI-assisted workflow for job discovery, contact research, venture mapping, application preparation, and outreach communications. Related venture-ops tooling produces briefings and reviewable drafts.

For this portfolio, application preparation is distinguished from verified application submission, and outreach drafting from confirmed sending. I have not established productivity, adoption, or response-rate improvements for this work. Its relevance is the operational workflow: research becomes an artifact that can be inspected before a consequential action.

Life Ops remains an integration effort

LifeOS / Life Ops is intended to consolidate personal tasks and venture commitments. Existing venture-ops code provides daily briefings, project tracking, and reminder/calendar integration paths; Venture Command provides structured commitment exports.

Unified ingestion, cross-source reconciliation, adaptive scheduling, and longitudinal personal intelligence remain development goals. They describe where the architecture is heading, rather than completed capabilities.

The engineering decisions I want to make inspectable

Across these systems, persistent state records what happened, versioned sources preserve what a plan relied on, validation constrains what can run, and approval binds a human decision to a specific artifact. Recovery behavior matters because an interrupted or uncertain action cannot always be safely repeated.

These are local applications and research implementations with explicit verification boundaries. The next useful evidence is repeated live workflow execution, recovery testing, and documented operating outcomes. I will extend the claims as those results exist.

Explore the implementations and related work through the project portfolio and the Aegis evaluation discussion.