The fastest, cheapest, most efficient open-source agent harness. Run more than 500 agents on a $10 VPS.
registry
User-facing agent registry. Owns the agent_registry RPC namespace: the
merged list of shipped default agents and user-authored custom agents, their
enable/disable state, and their tool visibility policy. This is a config
layer over the harness: the tool-calling loop and prompt/runtime
implementation live in agent/harness, and the definitions
the harness runs come from agents/ below.
Files
types.rs: wire types.AgentRegistryConfig(the persistedentries: Vec<AgentRegistryEntry>, stored atConfig.agent_registry),AgentRegistryEntry(id, name, description,source: AgentRegistrySource,DefaultvsCustom, enabled, model, optional system prompt,tool_allowlist/tool_denylist,subagents: AgentSubagentPolicyallowlist, tags, free-form metadata),AgentRegistryPatch(partial update),AgentToolInfo(tool-picker row).AgentSubagentPolicyalso deserializes from a legacy bare list of ids.AgentRegistryEntry::validateenforces non-blank id/name/description and an ASCII[A-Za-z0-9_-]id charset for both the entry and its subagent allowlist.defaults.rs:default_agents()loads the built-ins fromagents::load_builtinsand maps eachAgentDefinitionto aDefault-sourced registry entry (tier becomes the single tag,ToolScope::Wildcardrenders as["*"]).definition_from_registry_entryis the inverse: it synthesizes anAgentDefinitionfrom a custom entry (one with no shipped harness definition) so a custom agent runs through the sameAgent::from_config_for_agentfactory path as a built-in, with a real tool belt instead of a persona-only completion. Every harness-only field (omit_*, temperature, iteration policy, sandbox mode, tier) takes the harness's own safe worker default; see the function's doc comment for the exact mapping and why an emptytool_allowlistmust stayNamed(vec![])rather than collapsing toWildcard.ops.rs: config-backed CRUD.list_agents,get_agent,upsert_custom_agent(rejects ids that collide with a default),update_agent(copies the default into config on first edit; refuses to disableorchestrator),set_agent_enabled,remove_agent,merge_entries(defaults overlaid with persisted config entries, default-first),available_tools,find_custom_in_config(synchronous, matches only enabledCustomentries, used by the agent factory on a harness-registry miss). It reads and writes throughconfig::rpc::load_config_with_timeoutandConfig::save. Tool listing (available_tools) builds anorchestratorsession and reads its durable tool registry (durable_tool_specs_arc()), so it lists every registered tool, includingDeferredones the orchestrator reaches throughtool_searchand packed ones behinduse_skill, not only the schemas it advertises per turn.schemas.rs:ControllerSchema/RegisteredControllerdefinitions for namespaceagent_registry:list,get,available_tools,create_custom,upsert_custom,update,set_enabled,remove. Registered throughall_agent_registry_registered_controllersincore/all.rs.rpc.rs: request/response payload types and the*_rpchandler functions thatschemas.rswires up, delegating intoops.rs.tools.rs: backwards-compatible re-export ofagent::orchestration::tools::*; nothing in the tree imports it any more.agents/: the built-in agent archetypes and their loader.
agents/
Each built-in agent owns a subfolder with an agent.toml (id, when_to_use,
model, tool scope, sandbox mode, iteration cap, tier, omit_* flags, parsed
through tinyagents_definition::parse_str into AgentDefinition), and a
prompt.rs that exposes pub fn build(&PromptContext) -> anyhow::Result<String>.
Most builders include their archetype from a sibling prompt.md; the two flow
builders receive their archetype text from TinyFlows Copilot through an
explicit prompt context value. Builders append runtime-dependent sections such
as the rendered tool list and workspace to that archetype.
Every archetype uses the shared sub-agent turn graph.
agents/loader.rs owns the BUILTINS slice and
load_builtins, which parses each agent.toml, sets the dynamic-prompt marker
and stores its prompt callback in AgentDefinition.extensions.prompt_builder,
stamps DefinitionSource::Builtin, and checks that the folder id matches the
TOML id. It then calls DefinitionSet::validate_tier_hierarchy: a worker
may not list any agent-id subagent, and chat -> chat / reasoning -> reasoning
delegation is rejected (the pair rule is validate_tier_transition in the
owner definition crate; unknown subagent ids are tolerated here).
BUILTINS is also the registration point for archetypes that live with
other domains: skill_setup
(skills/catalog/agent/, feature skills), and
workflow_builder and flow_discovery (agents/, feature flows).
Workspace-level overrides (<workspace_dir>/agents/*.toml, with a
~/.openhuman/agents/ fallback) are loaded separately by
tinyagents_definition::DefinitionSet::load_dirs and replace built-ins on id collision;
AgentDefinitionRegistry::load re-runs the owner validation after that
merge.
The shipped archetypes in this directory:
| Archetype | Role |
|---|---|
image_agent |
Image generation/edit specialist |
morning_briefing |
Proactive scheduled daily summary (tasks, calendar, email, skills) on a named, read-only tool belt |
orchestrator |
Default user-facing chat-tier agent; direct-first, delegates only when it materially helps |
presentation_agent (feature documents) |
Builds decks from evidence; owns grounding/citations/image verification |
summarizer |
Runtime-dispatched only: compresses oversized tool results for the orchestrator |
task_manager_agent |
Task-source/workflow/artifact specialist: proactive feeds, workflow bundles, artifacts |
trigger_reactor |
One or two tool calls in direct reaction to an external trigger, no planning |
trigger_triage |
Classifies an external trigger into drop/acknowledge/react/escalate; never acts |
video_agent |
Video generation/animation specialist |
vision_agent |
Read-only image understanding: describe, OCR, locate UI elements |
The orchestrator's chat delegates (its [subagents] allowlist) are
task_manager_agent, vision_agent,
image_agent, video_agent, presentation_agent, skill_setup,
workflow_builder and flow_discovery. The retired workflow-run planner and
critic archetypes are no longer built-ins. Everything else is runtime-only.
The single-belt specialists that used to sit here (code_executor,
crypto_agent, settings_agent, scheduler_agent, mcp_agent,
tools_agent, help, tool_maker, skill_creator, context_scout,
integrations_agent, skill_executor, researcher) were removed. Their
playbooks are now inline skill guides on tool packs (coding, web3,
system, scheduling, docs, mcp) that the orchestrator loads with
use_skill, and their tools are Deferred, so tool_search finds them
too; see tools/toolpacks/README.md.
Running an installed skill is the orchestrator's own run_workflow.
presentation_agent stays compiled but
builtin_enabled drops it from load_builtins without documents, in
lockstep with its generate_presentation tool.
The orchestrator picks tools at request time through the shared tool-search
ranker (Jev when a decision-model credential is installed, BM25 otherwise);
see tinyagents/README.md and
the Jev page for how that
ranking works.
Called by
core/all.rs: registers theagent_registrycontrollers underDomainGroup::Agent.config/schema/:Config.agent_registry: AgentRegistryConfigis the persisted store everyops.rsfunction reads and writes.host_definition.rs: seeds the process-globalAgentDefinitionRegistryfromagents::load_builtins(); TinyAgents validates the merged hierarchy when workspace overrides load.agent/session_host/builder/factory.rs:Agent::from_config_for_agentfalls back tofind_custom_in_config+definition_from_registry_entrywhen an id is not in the harness registry.agent_runs/schemas.rs:agent.registry_snapshotenumerateload_builtins().flows/(ops/inference_readiness.rs,ops/builder_gates.rs,builder_tools/kind_reads.rs,tinyflows/caps/agent.rs): resolve a flowagentnode'sagent_refthroughlist_agents/get_agent/find_custom_in_config.
Tests
defaults_tests.rs, ops_tests.rs, schemas_tests.rs, types_tests.rs;
under agents/: loader_tests.rs (+ loader_tests_orchestrator_tier_tests.rs,
loader_tests_builtin_registration_tests.rs, loader_tests_specialist_agents_tests.rs)
and a prompt_tests.rs beside almost every archetype's prompt.rs.