A framework forproductionAI agents.
Zil is an open methodology for building, packaging, and operating AI agents at enterprise scale. Seven pillars, one signed artifact, deployable anywhere.
Production AI agents have outgrown the practices we use to ship them.
A short-turn chatbot fails in predictable ways. A multi-step agent that reasons, uses tools, maintains memory, and hands off to other agents fails in ways that compound — a bad reasoning step leads to a wrong tool call, which writes incorrect data, which poisons the agent's memory for the next session.
Traditional production-readiness checklists were built for the first generation of AI features. They do not cover what production agents actually need: lifecycle governance, agent-specific security, memory as a system of action, long-running execution, multi-agent coordination, and packaging that survives the next vendor migration.
46% of organizations cite integration with existing systems as their primary deployment challenge.67% aim to avoid high dependency on a single AI provider.35% admit they could not immediately disable a rogue AI agent.
Seven pillars. One discipline.
Each pillar is independently assessable, independently actionable, and designed to integrate with existing security, governance, and engineering practices. Together they describe what a mature agent operation looks like and how to get there.
- 01
Governance & Lifecycle
Who owns each agent, how it changes, when it retires, and how humans stay in the loop. Every agent has a registry entry, a named owner, an approval workflow, and a defined oversight UX.
- 02
Security & Adversarial Hardness
Prompt injection, tool-use abuse, memory poisoning, A2A spoofing, and supply chain risk. A repeatable threat model and red-team playbook for the agent attack surface.
- 03
Data & Memory Protection
Memory is data. Data has laws. Right-to-forget cascades, residency mapping, semantic context as systems of action — the agent-scale data layer treated as infrastructure, not afterthought.
- 04
Observability & Reliability
OpenTelemetry agent spans, reasoning traces, drift detection, crash recovery, idempotency. Long-running agents treated as a distinct architectural pattern with its own engineering practices.
- 05
Evaluation & Quality
Pre-deployment suites, multi-turn evaluation, tool-use correctness, planning quality, eval-in-production. Every promotion gated. Every production signal becomes a future test case.
- 06
Cost & Resource Governance
Token budgets at multiple granularities, model routing by task complexity, multi-provider fallback. The unit economics that determine whether agent programs survive scale.
- 07
Architecture & Packaging
Agents as portable, signed, versioned artifacts. Multi-agent orchestration as a first-class concern. The .zil package format separates what ships from how it runs.
What makes .zil different.
Open
Composes with MCP, A2A, and emerging open standards from the Linux Foundation Agentic AI Foundation. Zil is the layer above — packaging, runtime, lifecycle.
Portable
Adapter pattern for every external dependency. Switch LLM providers, vector backends, or vendors with a configuration change. No code rewrite.
Compliance-grade
Designed for regulated industries from day one. Audit trails, data residency, right-to-forget cascades, signed artifacts, SLSA provenance — the discipline regulators expect.
Practitioner-led
Refined through real engagements. Every primitive in the framework solves a specific failure mode that practitioners have actually hit in production.
.zil — a signed, portable agent bundle.
Every Zil-conformant agent ships as a single signed archive containing its full declarative specification. Manifest, identity, skills, memory configuration, RAG snapshots, and evaluation suite — packaged together, versioned together, deployed anywhere.
# customer-support-agent-2.1.4.zil
apiVersion: zil/v1
kind: Agent
metadata:
name: customer-support-agent
version: 2.1.4
spec:
identity: ./identity/persona.md
adapters: ./adapters/
skills: ./skills/
memory: ./memory/backend.yaml
mcp_servers: ./mcp/
data: ./data/kb_snapshot.tar.gz
evals: ./evals/baseline.yaml
# attestations
signature: cosign
provenance: slsa-v1
sbom: cyclonedx-1.5From the team at FluentData.
Zil is the methodology FluentData uses to deploy production AI agents for our clients. We are publishing it openly because we believe production agent operations need a shared vocabulary, and because no single vendor's platform should be the default answer for an industry-wide question.
FluentData is a forward-deployed engineering firm working with channel partners and direct clients on agentic AI delivery. Zil is the connective tissue across our engagements — refined through real production work, published as a signal of how we think about the problem.