Agentic Orchestration is a YAML-driven multi-agent engine on CrewAI. This page summarizes what the tool can do end to end — from dynamic planning through catalogs, execution backends, memory, and the web UI.

Dynamic Planning

State a goal in plain language and let the LLM planner (LiteLLM, configurable via AGENTIC_PLANNER_MODEL) build a multi-step JSON execution plan automatically.

Single-pass --dynamic

The planner reads the agent catalog, MCP catalog, session history, and knowledge-base snippets, then emits a JSON plan: workflow name, steps, agent provider IDs, MCP IDs, and per-step task descriptions. The coordinator runs each step in order via the active execution backend.

python main.py --dynamic "Compare three observability stacks for Kubernetes"

Iterative --dynamic-iterative

One step per round. After each step completes, the planner sees prior outputs and may change agents, MCPs, or task wording before the next round.

python main.py --dynamic-iterative --dynamic-iterative-auto \
  "Investigate market entry options for a medtech device"

Use --dynamic-iterative-auto for controller-driven early stop or continuation (up to --dynamic-iterative-max-rounds).

Domain-aware provider suppression

When goal text lexically matches specialist planner_hint or good_for fields, general-purpose agents (general_purpose: true) are suppressed so domain specialists win. Tune with AGENTIC_DOMAIN_PROVIDER_MATCH_MIN or disable via AGENTIC_DISABLE_DOMAIN_PROVIDER_SUPPRESSION.

See Dynamic Planning for sessions, attachments, and answer cache.

182-Provider Agent Catalog

Agent templates live as one YAML file per provider under config/agent_providers/. The dynamic planner selects from this catalog; constrain choices with --dynamic-agent-provider-ids ID1,ID2.

182 shipped templates across six provider types: openai, anthropic, ollama, huggingface, vllm, and jetstream.

Each YAML declares:

  • id — used in plans and CLI filters
  • type and model
  • role, goal, backstory
  • planner_hint / good_for — steering text for the planner
  • hardware.architecture and optional min_vram_gb

Extend the catalog without forking via AGENTIC_EXTRA_AGENT_PROVIDERS_CATALOG_DIRS (YAML directories) or AGENTIC_EXTRA_AGENT_PROVIDERS_PATH (Python provider classes).

Provider Example IDs Notes
openai gpt_research, gpt_write, gpt_reason Requires OPENAI_API_KEY
anthropic claude_research, claude_write, claude_reason Requires ANTHROPIC_API_KEY
ollama ollama_llama3, ollama_qwen2_5, ollama_llama3_2_vision Local via OLLAMA_HOST
huggingface hf_meta_llama_3_1_8b, hf_meta_llama_3_1_70b Requires HF_TOKEN
vllm vllm_tpu_meta_llama_llama_3_1_8b_instruct Requires VLLM_BASE_URL
jetstream jetstream_tpu_meta_llama_meta_llama_3_70b_instruct Requires JETSTREAM_BASE_URL

Full listing: Agent Catalog.

Platform Agent Harness (v1.4.0)

Verify catalog agents in isolation with tiered probes:

Tier CLI --harness-tier Checks
L0 static YAML + credentials
L1 connectivity Backend reachability
L2 smoke Single-task kickoff + assertions
L3 capability L2 + LLM rubric 
python main.py --harness-batch --harness-tier static
python main.py --harness-agent gpt_research --harness-tier smoke

Shared profiles in config/agent_harnesses/. CI runs L0 on every PR. Details: Agent harness roadmap.

User agent harnesses (v1.5.0)

Domain scenario packs per agent_provider_id — your prompts, fixtures, and rubrics under harnesses/<id>/.

python main.py --example healthcare --harness-agent gpt_research
python main.py --harness-dir ./my-harnesses --user-harness-run-all

Healthcare ships a reference pack. Details: User agent harnesses.

7 MCP Integrations

MCP tool definitions live under config/mcp_providers/. The planner attaches MCPs per step based on goal text, available credentials, and (for Kubernetes) transport compatibility.

MCP Transport Good for
search_brave streamable_http Web search, timely facts
search_tavily streamable_http RAG-optimized research excerpts
search_exa stdio (npx) Semantic search, code examples
home_assistant streamable_http Smart home control
fetch_url stdio (python) Fetch and parse any URL
filesystem_local stdio (npx) Scoped file read/write
memory_knowledge_graph stdio (npx) Persistent entity memory

Credential-gated entries stay hidden until required env vars are set. Merge custom YAML via AGENTIC_EXTRA_MCP_PROVIDERS_PATH.

Details: MCP Catalog.

Agent Skills (v1.3.0)

Procedural playbooks live as YAML under config/agent_skills/. Unlike MCP integrations, skills do not add callable tools — they inject curated markdown instructions into task descriptions (default) or agent backstory, composable with MCP and any agent provider.

python main.py --dynamic "Cut a patch release following our semver process"
# Planner may attach release_process when goal text matches

python main.py --config config/workflows/workflow_agent_skills_smoke.yaml "skills smoke test"
Skill id Injection Purpose
echo_skill task Smoke test — injects SKILL_ECHO_OK marker
echo_backstory_skill backstory Backstory injection smoke test
release_process task Semver release checklist (RELEASING.md, scripts/release.py)
pr_review task Structured pull-request review checklist

Attachment semantics mirror MCP: workflow.skills, per-task task.skills, planner skill_ids, and [] to force none. Workers on subprocess/Kubernetes backends re-resolve skill content from StepSpec.skills + catalog path.

CLI: --agent-skills-catalog (default config/agent_skills). Extend via AGENTIC_EXTRA_AGENT_SKILLS_PATH.

Details: Agent Skills · Agent Skills Roadmap.

Hardware-Aware Routing

Before planning, the runtime filters the agent catalog by detected hardware. Each provider YAML may declare architecture (cpu, gpu, tpu) and optional min_vram_gb. Detection uses nvidia-smi when available; Ollama models can use built-in VRAM heuristics.

hardware:
  architecture: [cpu, gpu]
min_vram_gb: 8
Variable Purpose
AGENTIC_ASSUME_GPU Force GPU as available
AGENTIC_ASSUME_TPU Force TPU as available
AGENTIC_ASSUME_VRAM_GB Override detected VRAM
AGENTIC_AVAILABLE_ARCHITECTURES Manual cpu,gpu,tpu override
AGENTIC_DISABLE_HARDWARE_FILTER Never drop providers by VRAM

Three Execution Backends

The pluggable ExecutionBackend protocol runs the same planner output and step contracts three ways:

Backend Env value How steps run Best for
CrewAI in-process inprocess (default) Whole crew in one Python process Local dev
Subprocess subprocess One worker per step via --execute-step Step isolation
Kubernetes kubernetes K8s Job per step, PVC run store Production

Planner JSON, YAML catalogs, session/KB/learning hooks, and post-run QA are identical across backends — only step isolation changes.

Details: Execution Backends.

Sessions, Learning, and Knowledge Base

Sessions

Planner turn history is persisted as JSON under __orchestrator_sessions__/<slug>.json. Name a session with --orchestrator-session or AGENTIC_ORCHESTRATOR_SESSION. Reset with --orchestrator-session-reset.

Knowledge base

When AGENTIC_KB=1, finalized outputs are indexed in SQLite FTS (__orchestrator_kb__/kb.sqlite3) and injected into future planner prompts as snippets.

Learning loop

AGENTIC_LEARNING=1 records per-provider success/failure stats and decision traces under __orchestrator_learning__/. Optional evaluator model via AGENTIC_LEARNING_EVAL=1. A learning summary feeds the planner on each run; web UI ratings are consumed on the next planner call. Ratings and KB rows use a combined attachment fingerprint (MCP + skill ids).

Answer cache

With AGENTIC_ANSWER_CACHE=1, asking the same goal in the same session returns the last finalized answer immediately and prompts to confirm a re-run.

Domain Verticals

Vertical overlays bundle domain orchestrator context, optional extra agent YAML, optional MCP YAML, and web start scripts — without forking core code. Load with --example <id>.

Vertical CLI Web UI port
Healthcare --example healthcare npm run start:healthcare → 3850
Logistics --example logistics npm run start:logistics → 3851
  • Healthcare — medtech evidence research, commercial brief generation, specialist agent templates.
  • Logistics — warehousing, WMS/ERP integration, labor planning, simulated WMS/ERP MCP fixtures.

Custom verticals: create examples/verticals/<id>/ with optional orchestrator-context.md, agent_providers/, and mcp_providers/. See Verticals.

WebSocket Web UI

The agentic-orchestration-web Node.js package streams run output over WebSocket, supports file uploads (attachment manifests for --dynamic-attachments), and exposes ratings that feed the learning loop.

cd agentic-orchestration-web
npm install
npm start
# http://127.0.0.1:3847

Vertical entry points: npm run start:healthcare, npm run start:logistics.

Static Workflows

For repeatable crews without LLM planning, define workflows under config/workflows/*.yaml.

# One-shot fixed workflow
python main.py --batch --config config/workflows/workflow_brainstorm.yaml

# Interactive loop on a fixed file
python main.py -i --config config/workflows/workflow_brainstorm.yaml

# Router: natural-language task → matching workflow
python main.py "Brainstorm taglines for a developer CLI"

Shipped workflows: default (workflow.yaml), brainstorm, web dev, healthcare commercial brief (router), MCP fetch/filesystem smoke tests, agent skills smoke (workflow_agent_skills_smoke.yaml).

Details: Workflows.