Proactive Agentic Orchestration

English | 简体中文

POrchestra enables SubAgents to proactively request task and resource adjustments during execution. The orchestrator revises the active agent's harness while preserving its local history, allowing execution and adaptation to form one continuous workflow.

The paper also introduces On-Policy Orchestration Distillation (OPOD), which turns execution-grounded feedback into process supervision for learning the orchestration policy.

Anonymous code repository

Existing re-orchestration methods typically wait for a SubAgent to finish or for a predefined trigger before adapting its configuration. POrchestra lets the executing SubAgent initiate communication whenever local observations reveal a need for adjustment.

Each SubAgent is configured with a subgoal, working context, tool set and execution budget. Its structured communication reports:

• State: progress, uncertainty and blockers.
• Request: the proposed task-scope or resource adjustment.
• Evidence: observations supporting the request.

The orchestrator can revise the active SubAgent, create a new SubAgent, terminate an execution path or complete the task. Applied revisions are recorded in the active agent's history so it can resume with its accumulated state. POrchestra uses one active SubAgent at a time.

OPOD combines feedback-guided distillation with outcome supervision:

1. Initialize the orchestrator through SFT on expert decisions.
2. Collect execution trajectories and attribute feedback to preceding delegation or revision decisions.
3. Convert attributable issues into natural-language improvement hints.
4. Sample actions from the current student and distill toward a frozen SFT teacher conditioned on the original context plus the hint, using full-vocabulary reverse KL.
5. Combine distillation with reward-weighted regression on decisions from successful trajectories.

The implemented objective described in Appendix E is L_OPOD = L_RWR + λ L_OPD, with λ = 1 as the default. The attribution model is the frozen base Qwen3.5-9B; the student starts from its SFT checkpoint. SubAgent models remain fixed.

porchestra/                 # MainAgent, SubAgents, communication and revision
  gaia2_agent/              # GAIA2 agents, prompts and mechanism ablations
aorchestra/                # Shared orchestration utilities and baseline
base/                      # Agent, model and memory infrastructure
benchmark/                 # GAIA, GAIA2 and SWE-bench adapters
config/example/            # Model and benchmark configuration templates
integrations/              # ARE compatibility patch
bench_porchestra_gaia.py
bench_gaia2.py
bench_porchestra_swebench.py
sitecustomize.py            # ARE startup registration

Run commands from the repository root using Python 3.13.

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
cp config/example/model_config.yaml config/model_config.yaml
mkdir -p config/benchmarks
cp config/example/benchmarks/*.yaml config/benchmarks/

Configure model endpoints and API keys through .env and config/model_config.yaml. GAIA web tools use SERPER_API_KEY and JINA_API_KEY. Multimodal tasks additionally require compatible image/audio models and relevant processing libraries. SWE-bench requires Docker and the official harness:

pip install -r requirements-swebench.txt

Download the datasets using the links below.

The GAIA2 evaluation and training pools are distinct: the paper uses the remaining 512 scenarios from the four categories for training. GAIA orchestrator training uses 2,339 TaskCraft tasks.

The example configurations run one task with concurrency one using gpt-4o. Configure the models and task limits in the benchmark YAML files. GAIA2 uses a separate ARE environment; follow GAIA2.md to install it.

# GAIA
python bench_porchestra_gaia.py --config config/benchmarks/porchestra_gaia.yaml

# GAIA2: install the ARE environment first (GAIA2.md)
workspace/are-env/bin/python bench_gaia2.py --config config/benchmarks/porchestra_gaia2.yaml --dry-run
workspace/are-env/bin/python bench_gaia2.py --config config/benchmarks/porchestra_gaia2.yaml

# SWE-bench Verified
python bench_porchestra_swebench.py --config config/benchmarks/porchestra_swebench.yaml

Results and trajectories are written to workspace/. GAIA and SWE-bench accept --tasks and --max_concurrency; GAIA2 accepts --limit and --concurrency.

GAIA and GAIA2 allow creation of up to 10 SubAgents. API identifiers are gemini-3-flash-preview, deepseek-v3.2 and deepseek-v4-flash; manuscript experiments use ChatAnywhere API access. For the paper experiments, use the corresponding task IDs, prompt variants, judge settings and execution budgets.

Prompt templates:

• GAIA: porchestra/prompts/main_agent.py and porchestra/subagent.py.
• GAIA2: porchestra/gaia2_agent/prompts.py and its communication modules.
• SWE-bench: porchestra/prompts/swebench_main_agent.py, swebench_subagent.py and swebench_mini_subagent.py.

Appendix F presents the core templates with placeholders for long runtime inputs. Configuration and environment variables select the runtime prompt variants.

The paper uses full-parameter SFT for three epochs, followed by one epoch of OPOD with LoRA (rank 32, alpha 64, dropout 0.05). Both use a learning rate of 1e-5, effective batch size 16 and a 32,768-token sequence setting. See Appendix E for sampling, loss normalization and evaluation details.

Release status: inference code for all three benchmarks is available. OPOD training code, data preparation, configurations and checkpoints will be added in a subsequent release.

The implementation builds on AOrchestra, and GAIA2 uses Meta Agents Research Environments. Source attribution and upstream licenses are preserved in THIRD_PARTY.md and licenses/.