Releases: pnnl/adept-agentic
Releases · pnnl/adept-agentic
Release list
v1.0.20260622-rc1
ADEPT (Agentic Discovery and Exploration Platform for Tools) v1.0.20260622-rc1
First public release. ADEPT is a production-ready agentic platform that connects
large language models to scientific computing infrastructure through a secure,
three-tier architecture with OpenAI-compatible APIs.
Platform architecture:
- Three-tier security model: Agent Gateway (authentication proxy with Keycloak
JWT/RBAC), Orchestration Service (LangGraph state machine with PostgreSQL
persistence), and stateless MCP Tool Servers - Dual API surface: OpenAI-compatible Response API (primary, streaming) and
Assistants API (legacy compatibility) — drop-in replacement for existing
OpenAI client code - Multi-agent orchestration: Router and Graph execution modes with dynamic
tool discovery, role personas with purpose-specific LLM routing via YAML
model catalog, and session-scoped state isolation - Agent-to-Agent (A2A) mesh: Cross-gateway delegation for federated
multi-site deployments via Gateway Registry - Provider-agnostic LLM layer: Azure OpenAI, AWS Bedrock, Ollama, and
Anthropic backends through LiteLLM with per-purpose model routing
Scientific computing capabilities:
- HPC integration: Agent-driven Slurm job submission with session-scoped
workspaces, SIF container execution, multi-agent LLM-as-judge validation - MCP tool ecosystem: 28 built-in tools spanning BLAST, UniProt, PubChem,
AlphaFold, web search, RAG (ChromaDB + PDF/CSV), code execution sandbox,
Nextflow pipelines, video transcription - Secure code execution: nsjail-isolated sandbox with Docker-in-Docker,
session-scoped file systems - File management: Upload, registration, presigned URL generation with
HMAC-SHA256 signing, batch processing with RAG index creation
Developer experience:
- CLI and SDK: Python SDK with device-flow authentication, interactive CLI
with background process management, A2A peer registration - IDE integration: VS Code MCP connector, Claude Code subagents,
29 registered development lifecycle skills - Deployment automation: 6-phase pipeline, Ansible playbooks, CDK/Pulumi
IaC, SSM credential injection - Observability: Langfuse tracing, Fluent Bit log aggregation, structured
cost tracking per LLM purpose
Security:
- Keycloak RBAC with group-based ACL tool provisioning
- 7-pass pre-release security scanner (secrets, PII, CVE, toxicity)
- Session isolation across all tiers (file, state, execution, network)
- HMAC-SHA256 presigned URLs with constant-time validation
License: BSD-2-Clause