Skip to main content

Secure agent architecture

MCP servers expose bounded tools. Agents combine a goal, model, memory, policy, and tools into a workflow. Dockprom-AI should begin with read-only diagnostic agents and add mutation only after approval, isolation, audit, and rollback controls have been proven.

  1. Observability Triage Agent: correlates alerts, metrics, logs, and traces.
  2. AI/LLM Performance Agent: diagnoses Ollama, token, context, queue, and GPU behavior.
  3. Runbook Recommendation Agent: retrieves cited operational guidance.
  4. AI Security Detection Agent: identifies injection, leakage, abuse, and anomalous tool activity.
  5. RAG Quality Agent: monitors retrieval relevance, coverage, and freshness.
  6. MCP Tool Policy Agent: makes deterministic allow, deny, or approval-required decisions.
  7. Incident Correlation Agent: produces a draft incident and evidence timeline.
  8. Remediation Planner Agent: proposes bounded, reversible steps without executing them.

The approval-gated Change Executor Agent should be last. It must never have a general-purpose shell or unrestricted Docker socket.

Harness flow​

Untrusted input
-> input classification and injection checks
-> bounded agent goal
-> read-only evidence collection
-> typed plan
-> deterministic OPA policy decision
-> human approval when required
-> narrowly scoped executor
-> independent verification
-> rollback or completion
-> immutable audit event and OpenTelemetry trace

OWASP-aligned controls​

  • Immutable agent goals and prompt/data separation.
  • Unique agent identity and least-privilege tool authorization.
  • Strict JSON Schema validation for every tool and inter-agent message.
  • Planner, policy, approval, executor, and verifier separation.
  • Tenant-isolated memory with provenance, classification, expiry, and quarantine.
  • Sandboxed non-root execution with restricted egress and short-lived credentials.
  • Signed/pinned agent, skill, MCP, model, and container dependencies.
  • Maximum steps, tools, tokens, retries, duration, and delegation depth.
  • Kill switches, circuit breakers, canaries, and tested rollback procedures.
  • Redacted OpenTelemetry traces and append-only structured audit events.
  • Human-facing evidence, uncertainty, blast radius, and rollback before approval.

The complete agent portfolio, secure skill contract, evaluation gates, and implementation phases are maintained in docs/AGENT_ROADMAP.md. Use the OWASP agent, MCP, and RAG controls as the mandatory design and release checklist.