swarms (@swarms_corp): Introducing Swarms v10: Async Sub-Agents, SkillOrchestra, and More

Introducing Swarms v10: Async Sub-Agents, SkillOrchestra, and More Over the past month and a half, the Swarms team has been heads down building toward the premier release of Swarms v10. What started as a series of targeted improvements quickly grew into one of the most comprehensive updates the framework has seen. Across dozens of commits and multiple pull requests, we rebuilt core systems, introduced powerful new orchestration primitives, and made the entire framework faster, smarter, and easier to work with. The sub-agent system received a full overhaul. Agents can now autonomously spawn sub-agents, assign them tasks, monitor their progress in real time, and cancel them on demand all backed by a persistent async task registry that handles retries and failure recovery without any manual wiring on your end. The `HierarchicalSwarm` took a major step forward with two standout additions. Worker agents can now execute in parallel, cutting down task completion time significantly on complex workflows. A built-in judge agent can also evaluate every worker's output after each cycle, scoring performance across five dimensions and surfacing actionable feedback so your swarm gets smarter with every run. We also implemented `SkillOrchestra`, a brand new routing primitive that lets you define agent skills explicitly and route tasks to the most qualified agent automatically based on what each agent knows how to do. Instead of hardcoding which agent handles what, `SkillOrchestra` makes skill-based delegation a first-class feature of the framework. On top of these headline features, the CLI was modernized from the ground up, a range of long-standing bugs were resolved, and internal defaults and cleanup work has made the entire codebase more stable and predictable. This changelog covers the most important features and improvements shipped during this period, organized by category. Async Sub-Agent Orchestration The sub-agent system received a major overhaul, making it significantly more powerful and easier to rely on for complex, multi-step workflows. Agents can now create and assign tasks to other agents autonomously meaning your orchestrator no longer needs manual wiring to delegate work. Async execution was introduced so sub-agents run in the background without blocking your main agent, and a full task registry backs every spawned sub-agent with automatic tracking, retries, and cancellation if something goes wrong. You can now check the status of any running sub-agent or cancel its tasks on demand, giving you full visibility and control over long-running workflows. Sub-agent output is also now printed in real time so you can see exactly what each agent is doing as it works. The entire system is cleaner and more stable under the hood, so you get faster, more reliable multi-agent pipelines with less setup and fewer surprises. Docs: https://docs.swarms.world/en/latest/swarms/examples/sub_agent_tutorial/ Github: https://github.com/kyegomez/swarms Hierarchical Swarm Improvements The `HierarchicalSwarm` now brings 2 powerful new capabilities to make your multi-agent pipelines faster and smarter. Parallel Agent Execution Judge Agent that evaluates worker agent outputs and provides feedback Enabling `parallel_execution=True` dispatches all worker agents at once using a thread pool, so instead of waiting on each agent one by one, they all run simultaneously dramatically cutting down total execution time on complex tasks. Setting `agent_as_judge=True` introduces a dedicated judge agent that scores every worker’s output after each cycle across five dimensions: task adherence, accuracy, depth, clarity, and swarm contribution. It produces a structured report with per-agent scores and tailored improvement suggestions, so you can quickly identify which agents are performing well and which need better instructions. Finally, `judge_agent_model_name` lets you choose exactly which model powers the judge use a lightweight model like `gpt-4o-mini` for fast, economical evaluations, or swap in a more capable model when precision matters most. Docs: https://docs.swarms.world/en/latest/swarms/structs/hierarchical_swarm/ Github: https://github.com/kyegomez/swarms SkillOrchestra Implementation SkillOrchestra is a new agent routing system, inspired by the research paper SkillOrchestra: Learning to Route Agents via Skill Transfer (https://arxiv.org/abs/2602.19672), that automatically sends each task to the most qualified agent based on skills rather than manual configuration. When a task comes in, an agent infers which skills it requires and how important each one is. Every agent in your pool has a competence score for each skill, and a weighted scoring formula ranks agents against the task's requirements in real time no agent calls needed for the scoring itself, just fast math. The top-scoring agent (or agents, if you set `top_k_agents > 1`) is selected and executes the task concurrently via a thread pool. Optionally, a learning mode evaluates output quality after execution and updates each agent's competence scores over time using exponential moving average, so the routing gets more accurate the more you use it. The result is a system where you define what your agents are good at, and `SkillOrchestra` handles the rest, there no hardcoded routing logic, no guesswork. Github: https://github.com/kyegomez/swarms Docs: https://docs.swarms.world/en/latest/swarms/structs/skill_orchestra/ Swarms CLI Overhaul The CLI received its most substantial overhaul to date: the old ASCII art banner has been replaced with a clean, compact header that displays your active model provider at a glance along with rotating startup tips and a simplified red-and-white color scheme. Help output has also been reworked; rather than rich-formatted tables, you now see a plain-text argparse-style reference that lists every command, option, and example in a format that is easier to scan and copy from. Throughout the CLI, unnecessary elements have been removed including the deprecated features command, redundant help subcommand, and unused CLI table utilities—yielding a leaner, more predictable interface. The swarms chat feature was fixed, and the CLI now runs with fewer required parameters, accelerating setup. Under the hood, the auto chat loop is now a persistent while-loop with exit-command handling, agent prompts receive a dynamically injected timestamp at runtime, and new bash command security guardrails block dangerous shell patterns including recursive deletion, fork bombs, and privilege escalation—before any command can reach your system. Docs: https://docs.swarms.world/en/latest/swarms/cli/cli_reference/ Improvements PDF Utility Removal — Removed pdf_to_text as a core utility from utils/__init__.py and data_to_text.py; inlined a local implementation directly in the mem0 example script. (2026-03-12) Agent Class Cleanup — Removed duplicated async sub-agent methods from the Agent class body, now superseded by registry integration. (2026-03-12) top_p Default Fix — Removed hardcoded top_p=None from auto_chat_agent and set Agent.top_p default to None to avoid unintended sampling parameter injection. (2026-03-12) Model & Version Update — Bumped swarms to 9.0.3 and updated default model references from gpt-4.1/claude-sonnet-4-5 to gpt-5.4/gpt-5.1 across example files and CLI. (2026-03-12) Async Sub-Agents Code Cleanup — Code cleanup and logic refactoring for async sub-agent execution. (2026-03-09) Hierarchical Swarm Docs Overhaul — Overhauled hierarchical_swarm.md with updated Mermaid diagram, new constructor parameter table, dedicated sections for all new features, arun() docs with FastAPI integration example, and updated best practices. (2026-03-05) CLI Cleanup — Removed deprecated features command, show_features function, redundant help subcommand, and dead CLI table utilities (create_command_table, create_detailed_command_table). (2026-03-03) Agent Defaults — Agent now defaults to gpt-4.1, uses styled formatter console input prompts in interactive mode, and always returns the final summary from the autonomous loop. (2026-03-03) CLI Documentation — Updated CLI documentation to reflect recent CLI overhaul. (2026-03-03) CLI Parameters — Improved CLI with fewer parameters and fixed the swarms chat feature. (2026-02-02) Sub-Agent Tutorial Docs — Improved sub-agent tutorial documentation. (2026-02-02) Examples Section — Improved examples section with updated references. (2026-02-05) Bash Tool for `agent.auto` — Added bash tool support for agents running in auto mode. *(2026-02-02)* Auto-Save Examples — Added examples for auto-saving structures. *(2026-02-02)* Bash Command Security Guardrails — Added security validation to block dangerous shell patterns including recursive deletion, pipe-to-shell, disk writes, fork bombs, and privilege escalation; enforces a 512-character command length limit. *(2026-03-03)* Bug Fixes Fixed README and auto-chat issue on CLI. (2026-03-11) Welcome Workflow — Fixed unexpected inputs in welcome workflow. (2026-02-28) LLMCouncil — Fixed incompatibility between LLMCouncil, SwarmRouter, and the API server. (2026-02-24) Agent Identity — Fixed system prompt injection for agent identity when no system prompt is provided. (2026-02-17) DebateWithJudge — Fixed API field for DebateWithJudge to work correctly with the API server. (2026-02-15) RoundRobin — Fixed RoundRobin swarm; removed useless parameter. (2026-02-15) LiteLLM Empty System Prompt Fix — Fixed Anthropic API rejections by stripping whitespace-only system prompt blocks and normalizing orchestrator-style System:/Human: prompts before message construction. (2026-03-12) Fixed typo: seperatedly → separately. (2026-02-10) Conclusion This release represents one of the most significant leaps forward in the Swarms framework to date. Over the past six weeks, nearly every major system sub-agent orchestration, the hierarchical swarm, the CLI, and core agent infrastructure has been meaningfully improved. Multi-agent workflows are faster with parallel execution, smarter with the judge agent, and more autonomous than ever with a fully async sub-agent registry that handles task tracking, retries, and cancellation out of the box. The CLI is cleaner, the defaults are more sensible, and a wide range of bugs that affected real-world usage have been resolved. I f you are building simple pipelines or complex multi-layered multi-agent systems, the framework is more capable, more reliable, and easier to work with than it has ever been. We are grateful for every contributor, bug reporter, and community member who helped shape this release and this is only the beginning of what is planned for v10. Get Started with Swarms Build, deploy, and scale your agents with Swarms now: Github: https://github.com/kyegomez/swarms Docs: docs.swarms.world Discord: https://discord.gg/2bZ37UmP2a