AgenticOrgLab/ExecOrchestra
AI EngineeringPython framework that turns tools, models, skills and agents into executable plans with three levels of model control.
What it does
ExecOrchestra is a Python orchestration framework for complex tasks that need more than one kind of resource. A plan decides which subtasks are handled by fixed tool calls and which need a model or an agent. It offers three execution modes:
- Direct: the plan fixes the operations and inputs; tools run without a model call in between.
- Localized: a model performs a bounded job, such as writing a short script, while tools handle reading, writing and computation.
- Extended: a model or agent can inspect a tool response and choose the next action within allowed tools and session limits.
Runs produce intermediate artifacts, a readable progress.log, detailed pipeline.log, and a delivery manifest, so developers can follow progress and diagnose failures. Benchmark entry points exist for GAIA and Terminal-Bench (benchmark docs).
Quick start
Requirements: Linux x86-64, Python 3.11+, Git, Conda, a running Docker Engine accessible to the current user, an OpenAI-compatible LLM endpoint, and an embedding service. The default embedding model is Qwen3-Embedding-4B.
-
Clone and install:
git clone https://github.com/AgenticOrgLab/ExecOrchestra.git cd ExecOrchestra/execorchestra conda create -n execorchestra python=3.11 -y conda activate execorchestra python -m pip install torch==2.11.0 --index-url https://download.pytorch.org/whl/cpu python -m pip install -r requirements.txt python -m scripts.prepare_release init test -f .env || cp .env.example .env -
Build the Docker runtime:
docker build -f sgar_mvp/docker/Dockerfile.runtime -t sgar-runtime:rc1 . python -m sgar_mvp.scripts.seal_rc1_runtime -
Set up embeddings. For local deployment, create a second environment, install
requirements-embedding.txt, and download the pinned model and tokenizer withpython -m scripts.download_embedding. Alternatively, use an existing endpoint that serves the same Qwen3-Embedding-4B revision and returns normalized 2,560-dimensional vectors; then download only the tokenizer (--tokenizer-only) and pointruntime.api_endpoint_urlandruntime.api_model_idinsgar_mvp/config/embedding_release.jsonat that service. -
Edit
.envwith your ownLLM_BASE_URL,LLM_API_KEY, andSGAR_EMBEDDING_API_KEY, then load it into the shell (set -a; source .env; set +a). -
If using local embeddings, start the embedding server in a second terminal (
python -m scripts.embedding_server, defaulthttp://127.0.0.1:8001/v1/embeddings). Then prepare resources:python -m scripts.prepare_release all --allow-unavailableThis checks models, roles, tools and agents, builds the resource index and verifies configuration. It makes live API and tool requests and may incur charges.
-
Run a task by writing the request to a UTF-8 text file inside an input directory, then:
bash scripts/run.sh \ --query-file "<task-file>" \ --public-input-root "<input-directory>" \ --run-dir "<new-run-directory>"
Results land in delivery_manifest.json (delivery_locator) and the status in run_manifest.json. An XBench DeepSearch example is available via bash scripts/run_xbench_demo.sh, with example details. For setup troubleshooting, see the installation guide, environment setup and model configuration.
Constraints
- Only Linux x86-64 is documented; the install path pins CPU-only PyTorch 2.11.0 and requires Conda and Docker.
- A compatible LLM gateway and an embedding service are mandatory inputs; the embedding endpoint must match the configured model revision and vector shape.
- Tool network access is disabled by default and must be enabled per run with
--network-policy declaredafter search/browser services are configured. - Preparing resources contacts live APIs and can generate usage costs;
--allow-unavailabletolerates missing candidates instead of failing. - No license is declared in the repository metadata, so redistribution and commercial use terms are unclear.
HotRepo's take
The project entered HotRepo's new-project Python list on 2026-11-16 with 106 stars gained over a roughly 24-hour window, and the sampled star count rose from 108 to 214 between 2026-11-15 and 2026-11-16. Beyond that pick-up, the record shows no GitHub daily rank or discussion threads, and the framework's main published companion is the CoEvoke dataset of multi-agent trajectories. The notable technical claim is the three execution modes, which let a plan spend model calls only where generation is needed; the install documentation is concrete, but the license gap is worth checking before adopting it in a product.
Written by AI from the project's README, updated . The project's own docs are the source of truth.