DevOps & AI agents glossary
Reference definitions for architecture patterns, protocol specifications, and operational concepts across the TalkOps multi-agent ecosystem and cloud-native infrastructure.
Quick index: A · B · C · D · E · G · H · I · J · K · L · M · O · P · R · S · T · W
A
A2A Protocol
Agent-to-Agent (A2A) Protocol. An open communication standard for structured messaging, capability discovery (Agent Cards), and state synchronization across distributed multi-agent networks using JSON-RPC 2.0 over HTTPS. See the A2A Protocol Reference.
A2UI Protocol
Agent-to-User Interface (A2UI) Protocol. A declarative streaming protocol that allows AI agents to render interactive user interfaces, status cards, and approval forms in client applications using schema-validated JSON rather than executable HTML/JS. See the A2UI Protocol Reference.
Agent
Autonomous Software Agent. A goal-driven system capable of inspecting infrastructure state, formulating multi-step execution plans, invoking external tools via MCP servers, and verifying outcomes within explicit approval boundaries. See the TalkOps Overview.
Agent Plugins (Agent Plugins 1.0.0)
Portable Multi-Component Plugin Standard. An open packaging specification supported across TalkOps, OpsCode, Claude Code, and OpenAI Codex that bundles subagents (agents/*.md), progressive-disclosure skills (skills/*/SKILL.md), slash commands (commands/*.md), MCP server configurations (.mcp.json), and lifecycle hooks (hooks/hooks.json). See the TalkOps DevOps Plugins Marketplace.
AGENTS.md
Persistent Project & Institutional Memory. A Markdown file loaded automatically by TalkOps and OpsCode to maintain repository conventions, architecture invariants, cluster naming rules, and team runbook standards across sessions. See Memory & Skills.
Alertmanager
Prometheus Alertmanager. The open-source routing, grouping, and deduplication engine for metric alerts. The Alertmanager MCP Server gives agents 14 tools to inspect firing alerts, create maintenance silences, and audit routing trees.
Approval Gate
Human-in-the-Loop Checkpoint. A deterministic governance boundary where an agent pauses execution before performing a mutating, destructive, or privileged operation, presents a syntax-highlighted diff or plan, and waits for explicit human confirmation.
ArgoCD
Declarative GitOps Continuous Delivery Engine. A Kubernetes controller that reconciles live cluster state against Git repositories. The ArgoCD MCP Server exposes 29 tools for inspecting sync status, reviewing manifest drift, and triggering rollouts.
Argo Rollouts
Kubernetes Progressive Delivery Controller. An advanced deployment controller supporting canary releases, blue-green promotions, and automated metric analysis. The Argo Rollouts MCP Server lets agents manage canary weight increments and rollbacks.
B
Blue-Green Deployment
Zero-Downtime Release Strategy. A release technique that runs two identical environments (active and preview) and switches live ingress traffic atomically once smoke tests and readiness probes succeed.
BranchMemoryStore
Isolated Subagent Memory Partition. A TalkOps memory architecture where verbose intermediate outputs—such as raw terraform plan logs or multi-file linter diagnostics—remain inside a subagent's private memory branch, returning only concise, validated artifacts to the supervisor.
C
Canary Deployment
Incremental Traffic Shifting Strategy. A progressive delivery pattern that routes a small percentage of live traffic (for example, 5% → 25% → 50%) to a new replica set while evaluating error-rate and latency SLOs before full promotion.
Central Server
Multi-Tenant LangGraph Runtime. The centralized execution mode (talkops serve) that hosts the TalkOps agent runtime, persists threads and long-term memory in PostgreSQL, ingests production webhooks, and serves connected CLI/TUI clients. See Central Server & Multi-Client Architecture.
Coordinator
Domain Task Coordinator. The mid-tier orchestration node in a hierarchical agent system that receives decomposed objectives from a supervisor and coordinates domain-specific subagents.
D
Deep Agent Pattern
Hierarchical Multi-Agent Architecture. An agent design pattern built on stateful graphs that combines explicit task planning, context-isolated subagents, persistent memory (AGENTS.md and skills), and deterministic tool guardrails.
Diataxis Framework
Documentation Architecture Standard. A technical documentation structure organizing content into overviews, quickstarts, task guides, and technical reference pages.
E
Error Budget
Reliability Allowance. The permissible amount of downtime or failed requests within an SLO window (for example, 0.05% for a 99.95% availability target). TalkOps SRE workflows inspect error budget burn rates before approving production rollouts.
G
GitOps
Git-Driven Operations. The practice of treating version-controlled Git repositories as the single source of truth for infrastructure and application state, applying changes via reviewed pull requests and automated reconciliation controllers.
Goal & Rubric Mode (--goal / /goal)
Autonomous Outcome-Driven Convergence Loop. A TalkOps execution mode where the agent evaluates its own output against explicit acceptance criteria or multi-criterion rubrics after each turn, iterating until all criteria pass or the turn budget is reached. See Goals & Rubrics.
Guardrails
Deterministic Safety Boundaries. Schema, regex, AST, and policy checks enforced before and after LLM tool invocations to block destructive commands, redact secrets, and enforce least-privilege access.
H
Helm
Kubernetes Package Manager. The standard packaging format (Chart.yaml, values.yaml, and Go templates) for Kubernetes applications. The Helm MCP Server provides 18 tools for chart discovery, template rendering, dry-run validation, and release rollbacks.
HITL (Human-in-the-Loop)
Explicit Operator Authorization. An execution model ensuring agents cannot mutate live clusters, apply Terraform plans, or delete resources without an authenticated human decision.
Hooks (PreToolUse / PostToolUse)
Deterministic Lifecycle Event Scripts. External shell or Python scripts executed outside the LLM context window on SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, and webhook.* events to enforce policy-as-code, block destructive mutations, redact secrets, and emit audit logs. See Lifecycle Hooks.
I
Infrastructure as Code (IaC)
Declarative Cloud Provisioning. Provisioning and managing VPCs, Kubernetes clusters, IAM roles, and managed databases through version-controlled code (Terraform, OpenTofu, CloudFormation) rather than manual console clicks.
Intent Recognition
Goal Classification. The initial supervisor reasoning step that classifies a user's prompt, extracts target environment parameters, and routes execution to the appropriate specialist subagent.
J
Jev Smart Mode
Dual-Engine <70ms Router & <100ms Safety Classifier. TalkOps's intelligent runtime mode (--smart / --approval-mode smart) powered by the TypeSafe AI Jev engine. It routes prompts across three model tiers (fast, standard, powerful) and reasoning levels in <70ms (reducing token costs by 80–90% on reflex queries) while scoring every tool call for mutation, blast radius (0.0–2.0), and intent alignment in <100ms with zero LLM prompt token overhead. See Approval Modes & Jev Smart Mode.
K
Kubernetes Agent (k8s-autopilot)
Autonomous Multi-Domain Kubernetes Agent. A LangGraph supervisor orchestrating 3 sub-supervisors, 13 subagents, and 8 MCP servers across Helm charts, cluster diagnostics, GitOps, progressive delivery, ingress routing, and observability. See the Kubernetes Agent Overview.
L
LangGraph
Stateful Multi-Agent Workflow Engine. The graph orchestration framework used by TalkOps to coordinate supervisor routing, cyclic validation loops, durable checkpoints, and interrupt() approval gates. See the LangGraph Reference.
Loki
Log Aggregation System. Grafana's label-indexed log storage engine. The Loki MCP Server provides 9 tools for discovery-first LogQL construction, log pattern extraction, and byte-scan cost guardrails.
M
MCP (Model Context Protocol)
Open AI Tool & Context Standard. A protocol that standardizes how AI agents discover, validate, and invoke external tools and read-only resources over stdio or streamable-http transports. See the MCP Ecosystem Overview.
MCP Resource
Read-Only Context URI. A structured, side-effect-free data endpoint exposed by an MCP server (such as loki://labels or argocd://applications) that agents read to gather environment context.
MCP Server
Tool Integration Gateway. A service implementing the Model Context Protocol to expose domain-specific DevOps tools (Helm, Prometheus, Terraform, ArgoCD) with strict JSON Schema validation and dry-run defaults.
MCP Tool
Executable Agent Function. A typed operation exposed by an MCP server (such as helm_upgrade or prometheus_query_range) with validated input parameters and structured JSON responses.
O
OpenTelemetry (OTel)
Vendor-Neutral Telemetry Framework. The CNCF standard for generating, collecting, processing, and exporting traces, metrics, and logs. The OpenTelemetry MCP Server provides 19 tools for Collector pipeline synthesis and auto-instrumentation.
OpsCode
Terminal-Native DevOps Coding Agent. An open-source, shell-first AI coding agent (ops) featuring 6 specialized subagents, 4-tier command safety classification, autonomous goal rubrics, and remote SSH/Docker/Kubernetes sandboxes. See the OpsCode Overview.
P
Progressive Delivery
Controlled Release Engineering. Deployment practices—including canary rollouts, blue-green switches, traffic mirroring, and automated metric analysis—that limit the blast radius of new releases.
Prometheus
Cloud-Native Time-Series Monitoring. The CNCF standard for metric collection and alerting. The Prometheus MCP Server exposes 28 tools for PromQL execution, exporter management, and TSDB cardinality analysis.
PromQL
Prometheus Query Language. The functional query language used to aggregate time-series metrics, calculate request rates (rate()), and compute latency quantiles (histogram_quantile()).
PSS Restricted
Kubernetes Pod Security Standard (Restricted). The strictest built-in Kubernetes pod hardening profile, requiring non-root containers (runAsNonRoot: true), read-only root filesystems, dropped Linux capabilities (drop: ["ALL"]), and RuntimeDefault seccomp profiles.
R
Remote Sandbox
Isolated Tool Execution Environment. A containerized or cloud-hosted execution target (Docker, Kubernetes pod, Modal, Daytona, or Runloop) where TalkOps executes shell commands and file edits via the sandbox-as-tool pattern to contain blast radius. See Remote Sandboxes.
Rollout
Managed Workload Transition. The progressive replacement of pod replicas across a Kubernetes Deployment or Argo Rollout resource while verifying readiness and metric health.
S
Server Addon Microkernel
Pluggable talkops serve Extension Architecture. The modular server architecture (ServerAddon ABC, AgentRunGateway, and HookEventBus) that powers built-in (core-auth, client-gateway, webhooks) and custom FastAPI routes, background workers, and ChatOps integrations in TalkOps Server Mode. See Server Addons.
SKILL.md (Agent Skills)
Progressive-Disclosure Operational Runbooks. Modular domain playbooks and cryptographically verified helper scripts loaded on demand when an agent matches a task domain, keeping the base context window lean. See Memory & Skills.
SLO (Service Level Objective)
Reliability Target. A quantitative threshold for service availability or latency (for example, 99.9% of requests completing in under 250ms over a rolling 30-day window).
State Machine
Deterministic Execution Graph. A workflow model consisting of typed state, processing nodes, and conditional transition edges.
Subagent
Domain-Specialized Worker. A focused agent with a restricted toolset and isolated context window responsible for completing a single domain task (such as writing a Helm template or auditing IAM policies).
Supervisor Agent
Root Orchestrator. The top-level agent that interprets user requests, decomposes multi-step plans, delegates work to subagents, and synthesizes the final response.
T
Tempo
Distributed Tracing Backend. Grafana's object-storage-backed trace engine. The Tempo MCP Server exposes 16 tools for TraceQL queries, critical-path waterfall compression, and RED metric extraction.
Terraform
Declarative Infrastructure as Code Engine. The standard tool for provisioning cloud resources across AWS, Azure, and GCP. The Terraform MCP Server provides provider documentation search and sandboxed init/validate/plan execution.
Traefik
Cloud-Native Edge Router. A Kubernetes ingress controller and reverse proxy. The Traefik MCP Server provides 11 tools for configuring IngressRoute CRDs, middleware chains, and weighted canary traffic splits.
W
Webhook Incident Triage
Autonomous Event-Driven Alert Investigation. The TalkOps workflow that ingests real-time alerts from 9 providers (PagerDuty, Alertmanager, Datadog, AWS SNS, CloudEvents, ServiceNow, Jenkins, GitHub, Azure Monitor), deduplicates alert storms, runs a LangGraph root-cause investigation, and writes structured RCAs back to tickets, Slack, and PRs. See Webhooks & Event-Driven Automation.
Workflow
End-to-End Agent Execution Path. The complete lifecycle of a TalkOps task from intent classification and subagent delegation through sandbox validation, Human-in-the-Loop approval, and execution verification.