Production Governance for
Agentic AI
Runtime enforcement, evidence trails, and human controls so teams can ship autonomous workflows without making “I hope it behaves” the safety model.
Jesse C
Founder, Veilfire
Jesse C
Founder, Veilfire
I work at the intersection of security engineering, distributed systems, and applied AI. I build systems for teams that want the upside of autonomous agents without the quiet failure modes that show up later: drift, privilege creep, unsafe tool use, and compliance ambiguity.
My stance is simple: governance has to live at the action boundary, not as an after-the-fact dashboard and not as transcript archaeology. Enforcement + evidence + escalation are defaults. That’s how agentic AI becomes trustworthy in production.
Background
- Security + distributed systems background focused on production controls, not just prototypes.
- Applied AI / agent workflows where autonomy meets real operational constraints.
- Built the Veilfire governance stack: Ember, FireDeck, Lens, and Insight.
- Publishes practical field notes, threat modeling, drift, and runtime governance patterns.
How I build
Production governance means the safety model lives inside runtime execution, not a checklist. Policy is enforced where actions happen, evidence is generated as part of operation, and escalation exists before incidents do.
What you can hold me to
Explainable decisions, explicit control for high-risk actions, and privacy-preserving observability — defaults, not vibes.
Privacy
- Metadata-first observability by default.
- Raw prompts and sensitive content stay in your environment.
- Data minimization is a design constraint, not a feature toggle.
Security Stance
- Least privilege + bounded autonomy for every agent.
- Enforcement at tool / action boundaries.
- High-risk operations escalate to approval / HITL.
Evidence + Retention
- Evidence over transcript hoarding.
- Tamper-evident audit trails tied to policy versions.
- Retention follows risk + compliance intent.
Proof in public
Our values aren't a pitch deck — they ship as software. Two free, open tools that prove governance and capability belong together.
Pyromancer
An operator-first AI terminal
Veilfire's values made tangible. Not a demo or prototype — a production tool, free and hosted on GitHub, built on safe, secure, ethical AI to prove that governance and capability are not at odds.
Operator Control
Human-in-the-loop by default. Three execution modes from step-by-step approval to bounded autonomy — you set the boundaries, the agent stays inside them.
Privacy & Security
No telemetry, no cloud dependency. Secrets in your platform's secure vault, terminal output redacted before the AI sees it, tamper-evident audit trails.
Open & Auditable
Free on GitHub for macOS and Windows. Cryptographic audit logs make every AI action, permission decision, and command verifiable locally.
EmberSpark
Open agentic AI for everyone
The same Safe, Secure, Ethical AI principles behind the Veilfire platform, handed to the developer community as open source. Apache 2.0, Python, Docker-deployable — bounded autonomy and operator control as the default, not an enterprise-only privilege.
Bounded Autonomy by Default
Closed-by-default tool permissions, declarative I/O, and BudgetGuard caps on iterations, calls, and cost. Safety lives at the design layer, not as a bolt-on.
Local-First & Private
Run agents fully offline with Ollama. No telemetry, no cloud dependency, no data leaving your machine. Privacy is the default, not an upgrade tier.
Apache 2.0 Forever
Open by default, free for any use, hostable anywhere. EmberSpark stays open source — always.
Thinking in the open
Ongoing field notes on agent threat modeling, drift detection, and production governance. If you want the implementation thinking, this is where it lives.
The Glass Protocol
A seven-part series on the assurance gap: why agents break older assurance models, and what an architecture of assurance looks like when safe, secure and ethical are treated as outcomes rather than features.
The Eval Was the Attack
The OpenAI x Hugging Face incident, and what it shows about trusting evaluation pipelines you don't control.
Your Writing Style as a Cross-Internet Tracking Signal
Stylometry turns prose into an identifier. What that means for pseudonymity when everything you write is a fingerprint.
EmberSpark: Open-Source Autonomy Without Unlimited Agency
Why we built an open-source agent runtime where bounded autonomy is the default, not an afterthought.
The AI Perimeter Has Moved to the Action Boundary
Network firewalls don't catch what an agent does. The perimeter for AI is now the action boundary — every tool call, memory access, and outbound request.
Pyromancer: An Operator-First AI Terminal
Built on safe, secure, ethical AI. An AI terminal designed with operator control and governance at its core.
EU AI Act Compliance + AI Agents
What you need to know about AI agent compliance under the EU AI Act.
Human-in-the-Loop Is a Privacy Trap
Why HITL oversight defaults to full exposure and how graduated disclosure keeps surveillance out of safety workflows.
Threat Modeling AI Agents
A practical threat modeling walkthrough of a hotel customer support agent.
From Threat Model to Runtime Control
Continuing the hotel CS agent story — a hands-on demo of HITL escalation and policy blocking with Veilfire Lens.
Tracking Drift
How autonomy creep happens, and how to detect it.
Make agentic AI production-grade.
Put guardrails where they matter: runtime enforcement, audit-ready evidence, and human control for high-risk actions, all without capturing raw prompts.
Proof, not promises.