Concepts
The 7 Engines
V1 ships the four engines that get an agent running safely in production. The remaining three turn single agents into governed, multi-agent systems, and are gated behind apiVersion: vesper/v2so today's manifests keep working.
1. Agent Registry Engine
Available (V1)Declarative YAML manifests with schema validation and git-like versioning, stored in a local registry. Apply, list, inspect, and roll through versions of every agent.
2. Agent Lifecycle Engine
Available (V1)Executes agents across OpenAI, Anthropic, and Google. Runs the tool-calling loop, loads built-in and custom tools, and records every run.
3. Agent Memory Engine
Available (V1)Persistent, stateful memory with session and project scopes, automatic retention windows, and overflow truncation, so agents remember across runs.
4. FinOps Engine
Available (V1)Tracks the cost of every run, enforces maxCostPerRun hard caps, warns at alertAt thresholds, and keeps a full audit log. Cost governance as a first-class citizen.
5. Agent Communication Bus
Coming soonOrchestrate multi-agent fleets: task graphs with dependencies and message passing between agents (kind: AgentFleet). The step up from single agents to coordinated systems.
6. Quality & Eval Engine
Coming soonAttach eval datasets and pass thresholds to an agent, then gate deploys in CI so regressions never ship.
7. Security Engine
Coming soonGuardrails and policy-as-code enforced at runtime around agent inputs, outputs, and tool use.