Capabilities

Capabilities

This is the discovery index for TinyAgents: one page where an agent (or a human) can see every functionality the crate exposes and jump straight to the deep page for it. Each row is a capability, a one-line description of what it does, and a Where link to the wiki page (and section anchor) that documents it. Start here, find the surface you need, and branch out.

The crate is organized into five surfaces — Harness, Graph runtime, Registry, Expressive .rag, and REPL .ragsh — plus Providers at the leaves and testkit/conformance for tests.

Harness — model calls, tools, control

Capability	What it does	Where
Agent loop	Default model→tool→model loop with run identity + limits	Harness
Provider-neutral model calls	One ChatModel trait; ModelRequest/ModelResponse across hosts	Harness
Model resolution & selection	ModelRegistry resolves a call via ModelSelection/ModelHint/capabilities	Harness
Typed tools	Tool trait with JSON schemas and a ToolRegistry	Harness
Middleware hooks	before/after hooks around agent, model, and tool calls	Harness
Structured output	Extract a typed value via a provider-appropriate strategy	Harness
Streaming	Model-delta and event streaming through the loop	Harness
Usage & cost	Token accounting and pricing roll-ups across the run tree	Harness
Limits, retry, fallback	Fail-closed caps, backoff, model fallback, rate limiting	Harness
Response & prompt cache	Local response cache + provider prompt/KV-cache layout	Harness
Memory & embeddings	Thread memory, vector stores, retrievers	Harness
Sub-agents (agents as tools)	Wrap an agent as a tool; child runs one level deeper	Harness
Steering	Typed runtime control (pause/resume/redirect) of a running agent	Harness
Summarization	Context-window-aware transcript compaction	Harness
Events & run status	In-memory event spine + HarnessRunStatus	Harness
Durable observability	Journals, status stores, sinks, latency metrics	Harness
Testkit	Fakes, recorders, trajectory assertions	Harness

Graph runtime — durable typed workflows

Capability	What it does	Where
Nodes & edges	Named nodes wired by static/conditional edges over typed state	Graph Runtime
Dynamic routing	Command.goto + routing precedence resolve next targets	Graph Runtime
Send fan-out (map-reduce)	Schedule nodes with per-invocation args for the map step	Graph Runtime
Reducers & channels	Merge branch updates deterministically; channel-per-field state	Graph Runtime
Parallel fan-out	Run multi-node supersteps concurrently, folded in index order	Graph Runtime
Checkpoints & durability	Persist at superstep boundaries; in-memory/file/SQLite backends	Graph Runtime
Time travel	Read/fork/update state history against a checkpointer	Graph Runtime
Interrupts & resume	Human-in-the-loop pause and resume with a payload	Graph Runtime
Subgraphs	Embed a compiled graph as a node (graph-runs-graph)	Graph Runtime
Sub-agent nodes	Embed a harness agent as a graph node	Graph Runtime
Recursion policy	Bound depth, per-node visits, and total steps	Graph Runtime
Orchestration tools	Model-callable child-work supervision (orchestrate_*)	Graph Runtime
Streaming & events	GraphEvents + StreamMode projections	Graph Runtime
Durable graph observability	Journals, status store, latency metrics	Graph Runtime
Topology export	JSON/Mermaid export of a graph's structure	Graph Runtime
Graph testkit	Deterministic node doubles + graph assertions	Graph Runtime

Registry — named capability catalog

Capability	What it does	Where
Component identity	ComponentId/ComponentKind/ComponentMetadata describe capabilities	Registry
Capability registry	Register/resolve models, tools, graphs, agents by name	Registry
Bind .rag/.ragsh by name	Allow-list validation of agent-authored source before it runs	Registry
Model catalog	Offline pricing, context windows, capability facts	Registry

Language & REPL — authored and interactive workflows

Capability	What it does	Where
Expressive .rag	Declarative, side-effect-free blueprints that compile to the runtime	Expressive Language (.rag)
REPL .ragsh	Imperative, capability-bound interactive orchestration (RLM loop)	REPL Language (.ragsh)

Providers — model backends at the leaves

Capability	What it does	Where
Offline mock model	Deterministic, network-free default build	Providers
OpenAI + compatible hosts	One adapter for OpenAI, Anthropic, Ollama, DeepSeek, Groq, xAI, OpenRouter, Together, Mistral	Providers
List available models	OpenAiModel::list_models() — runtime model discovery via GET /models	Providers
Provider inference	Resolve a provider from a model string	Providers
Capability profiles	Reject impossible requests; pick capability-satisfying fallbacks	Providers

Testing — testkit & conformance

Capability	What it does	Where
Harness testkit	Scripted models, fake tools, trajectory asserts	Harness
Graph testkit	Node doubles + fluent graph assertions	Graph Runtime
Storage conformance	Reusable contracts for task stores and checkpointers	Graph Runtime

Recently added (SDK-gap hardening)

New capabilities that harden the SDK surface. Deep docs live on the linked pages.

Capability	What it does	Where
Model lifecycle gating	ModelRegistry::resolve skips ModelStatus::Retired; ModelSelection.allow_retired = true opts back in	Harness
orchestrate_list filters	Filter managed tasks by kind, created_after_ms, created_before_ms	Graph Runtime
Unknown-tool recovery	UnknownToolCall carries original arguments; UnknownToolPolicy{Fail,ReturnToolError,Rewrite}	Harness
Stable event ids	EventSink::with_stream_id; ids {stream_id}-evt-{offset} survive restarts	Harness
Workspace isolation	RunContext::with_workspace, prepare/cleanup_workspace, WorkspaceDescriptor::enforce, AgentEvent::Workspace*	Harness
Tool exposure auditing	AgentEvent::ToolsFiltered; ContextualToolSelectionMiddleware::inheriting(...)	Harness
Control outcomes	AgentEvent::ControlApplied, MiddlewareControl::{kind,precedence}, precedence-based request_control	Harness
Tool policy enforcement	ToolPolicyMiddleware::{require_sandbox, require_approval, enforce_result_bytes}	Harness
Model delta attribution	AgentEvent::ModelDelta now carries run_id	Harness
Thinking-token streaming	MessageDelta.reasoning, ModelDelta.reasoning, and Usage.reasoning_tokens keep thoughts separate from visible text	Harness
Budget reservation & cost	BudgetLimits.max_cached_input_tokens, AgentEvent::{BudgetReserved,BudgetReconciled}, preflight reservation	Harness
Parallel map/reduce controls	ParallelOptions::{with_item_timeout, with_total_timeout, with_cancellation}	Graph Runtime
Registry introspection	ComponentKind adds Middleware/Checkpointer/TaskStore/Listener; RegistrySnapshot.aliases; name-reuse diagnostic	Registry
Storage conformance	graph::testkit::conformance::{taskstore_concurrent_contract, checkpointer_concurrent_contract, taskstore_replay_contract}	Graph Runtime

Agent quick-start: pick your task

• Want to enforce tool safety (sandbox/approval/output caps)? → Tool policy enforcement
• Need a token/cost budget with preflight reservation? → Budget reservation & cost
• Got a model call for an unregistered tool? → Unknown-tool recovery
• Want to stop calling a deprecated model? → Model lifecycle gating
• Running agents in parallel with timeouts/cancellation? → Parallel map/reduce controls
• Need to isolate a run to a workspace? → Workspace isolation
• Want stable, replayable event ids? → Stable event ids
• Supervising child work from a model? → Orchestration tools
• Pausing for human input mid-graph? → Interrupts and resume
• Letting a model author its own workflow? → Bind .rag/.ragsh by name
• Verifying a run's shape in a test? → Harness testkit / Graph testkit

See also

• Home — orientation and the five-surface overview.
• Examples — runnable end-to-end demonstrations.
• Architecture — how the surfaces compose.