Concepts

The primitives every core-agent deployment touches, regardless of whether you’re running the CLI or embedding the library. Reference-shaped: each page is the depth material for one concept, not a tutorial.

ProvidersGemini (API + Vertex), Anthropic (first-party + Vertex), mock providers. Env vars, model IDs, auto-detection.

Permissionsask / allow / plan / yolo modes, pattern grammar, path + URL scopes, bash denylist, per-call vs. session grants.

Sessions and event logDurable SQLite/Postgres/MySQL-backed session storage, monotonic seq, Since(seq) replay, Watch(seq) live tail, crash-resume.

Context managementCompaction, task-boundary checkpoints, and agentic tool wrappers — the three mechanisms that keep long sessions alive past the context wall.

MCP serversDeclarative third-party tool integration via .agents/mcp.json — lifecycle, transports (stdio + Streamable HTTP), namespacing, gating.

SkillsClaude-compatible SKILL.md bundles, auto-discovery from project + user-global scopes, invocation on demand.

Built-in toolsFile, search, shell, data + network, planning, interactive-prompting tools. Every one gated, every one output-capped.

HooksLate-binding integration points — before-turn, after-tool, post-construct — for rate limits, custom budgets, external approvals.

Multi-session daemonOne server process serving many concurrent sessions with per-caller ACLs and per-session gate configuration.

OpenTelemetryOpt-in OTLP tracing over HTTP or gRPC — attribute cost, latency, and errors across model calls, tool invocations, and pruning passes.

Environment variables (env.yaml)Declare env vars the bundle expects, validate at boot, interpolate ${env:VAR} across AGENTS.md, skills, and mcp.json.