Key takeaways
- Hierarchical multi-agent system where a top-level planning agent coordinates specialized lower-level agents for both deep research and general-purpose tasks
- Built on the Autogenesis protocol — a self-evolution framework where agents can dynamically instantiate, refine, and version their own tools, prompts, and capabilities during execution
- Two protocol layers: RSPL (Resource Substrate) manages resources with explicit state and versioning; SEPL (Self Evolution) specifies the propose/assess/commit improvement loop with rollback
- 3.4k stars and 444 forks as of June 2026, MIT license. Still the most architecturally ambitious deep research agent, but commit cadence has slowed — last push May 4, 2026 — and independent community discussion is nearly absent
FAQ
What is DeepResearchAgent?
A hierarchical multi-agent system from Skywork AI that uses the Autogenesis self-evolution protocol. A planning agent coordinates specialized sub-agents that can dynamically create and refine their own tools and prompts during execution.
What is Autogenesis?
A self-evolution protocol with two layers: RSPL manages resources (prompts, tools, memory) with versioning and lifecycle; SEPL specifies how agents propose, assess, and commit improvements with auditable lineage and rollback.
Is DeepResearchAgent still maintained?
Partially. As of June 2026 the repository is not archived and was last pushed on May 4, 2026, but v2.0.0 ("self evolving") remains the only major release since the Autogenesis rewrite and activity has slowed to occasional maintenance commits.
Overview
DeepResearchAgent is Skywork AI's hierarchical multi-agent system designed for deep research and general-purpose task solving. Its distinguishing feature is the Autogenesis protocol — a self-evolution framework where agents can dynamically create, refine, and version their own resources during execution.
The architecture uses a top-level planning agent to coordinate specialized lower-level agents (domain agents, tool-calling agents), with an iterative Act-Observe-Optimize-Remember loop that enables agents to improve across runs.
Key stats (as of June 2026): 3,449 stars, 444 forks, MIT license, Python (99.8%), 10 open issues, last pushed May 4, 2026.
Status: Active but Slowing (June 2026)
The GitHub repository is not archived and received commits as recently as May 4, 2026. Since this profile's original publication in March 2026, stars have grown from ~3,241 to 3,449 — modest organic growth rather than breakout adoption.
The release history is thin: only two tagged releases exist — v1.0.0 ("pre version") and v2.0.0 ("self evolving"), the rewrite that introduced the self-evolution architecture — with nothing tagged since. Development since the v2.0.0 rewrite has been incremental maintenance rather than new capability.
Architecture: Autogenesis Protocol
Two protocol layers:
RSPL (Resource Substrate Protocol Layer):
- Models prompts, agents, tools, environments, and memory as protocol-registered resources
- Explicit state, lifecycle, and versioned interfaces
- Enables composable agent systems
SEPL (Self Evolution Protocol Layer):
- Closed-loop operator interface: propose, assess, commit improvements
- Auditable lineage and rollback capability
- Optimizers: reflection, GRPO, Reinforce++ methods
The iterative loop: Act (produce actions via LLM + tools) then Observe (capture outcomes and traces) then Optimize (update prompts/solutions using optimizer) then Remember (persist insights to memory).
Pricing & Licensing
| Tier | Price | Includes |
|---|---|---|
| Open Source | Free | Full framework, MIT license |
| LLM API costs | Variable | Bring-your-own keys (OpenRouter, GPT-4o, and other providers) |
Licensing: MIT — use commercially, modify freely.
Competitive Position
Strengths: Most architecturally ambitious deep research agent. Self-evolution is a genuine differentiator. Composable agent/tool/environment system. MIT license.
Cautions
- Complexity — many abstractions (RSPL, SEPL, optimizers, tracers, versioning) for a framework with a small contributor base
- Self-evolution benefits are theoretical until proven at scale — no published benchmark results for the Autogenesis architecture in the current README
- Slowing momentum — no release since v2.0.0 and commit activity tapered after early May 2026
- Thin independent validation — smaller community than simpler alternatives, and almost no third-party discussion (see below)
What Developers Say
No substantive independent developer commentary surfaced in research as of June 2026 — searches found no Hacker News threads, no Reddit discussions, and no hands-on third-party reviews of DeepResearchAgent specifically (coverage of Skywork's commercial skywork.ai product does not address this framework). The 3.4k stars and 444 forks are the strongest available adoption signal, but they are not corroborated by visible practitioner usage reports. Treat community validation as an open question.
Bottom Line
Not recommended for production deep-research workloads as of June 2026. The Autogenesis architecture remains the most intellectually interesting design in the autoresearch space, but slowing commits, a single major release, no published benchmarks for the current architecture, and near-zero independent practitioner feedback make it a research codebase to study, not a dependency to build on. Outlook: worth re-checking if Skywork ships a post-v2.0.0 release with benchmark results or the community discussion gap closes; otherwise expect it to remain a reference implementation for self-evolving agent ideas.
Research by Ry Walker Research