Skip to content

AstraDesk

Glossary

Core terms used across AstraDesk Framework 1.0 documentation.

  • ADLC - Agent Development Lifecycle: Plan → Build → Test/Optimize → Deploy → Monitor/Operate.

  • AstraOps - Observability + Evaluations tooling and dashboards that answer “Is it right?”.

  • AstraCatalog - Registry of agents, tools, prompts, datasets, risk posture, and certification artifacts.

  • Gateway - The control plane entrypoint for MCP tools and LLM routing with OPA, quotas, and audit.

  • MCP - Model Context Protocol for tools/resources/prompts discovered and invoked by agents.

  • OPA/Rego - Policy-as-code engine and language used for side-effect gating and data egress rules.

  • Acceptable Agency - Explicit authority limits for an agent; guards read|write|execute side effects.

  • Groundedness - Degree to which an answer is supported by provided context/evidence.

  • Containment - % of cases resolved without human handoff (business KPI).

  • SLO - Service Level Objective (e.g., latency p95 ≤ 8s, tool success ≥ 95%).

  • Champion–Challenger - Controlled promotion where a new version must outperform the current one on the same evals.

  • Schema Pinning - Requiring a hash of the tool schema with every call; the Gateway rejects mismatches.

  • Shadow/Canary - Running a challenger on mirrored traffic (shadow) or progressively shifting live traffic (canary).

  • Judge Kernel - Pluggable rubric/model that scores helpfulness, safety, groundedness for in-loop gating.

  • AstraGraph Memory - Planned hybrid vector + graph memory with temporal decay and provenance (v2.0).



Cross-References