ViewDAO
đź“Š Introducing Swarms v11: AdvisorSwarm, Swarms CLI, Heavy Swarm, and More
Over a five-week period, the Swarms team and community contributors shipped 120 commits across 35 pull requests, introducing 3 new swarm architectures, 7 bug fixes, 5 security patches, and removing over 13,000 lines of deprecated code to streamline and stabilize the core system.
New capabilities significantly expanded the framework’s coordination and reasoning depth. AdvisorSwarm introduced an executor–advisor pairing for dynamic guidance, PlannerWorkerSwarm added parallel task execution with dependency ordering and a judge cycle, and PlannerGeneratorEvaluator implemented a GAN-style generate–score–retry loop. HeavySwarm was also upgraded with Grok 4.20 Heavy support and expanded to a 16-agent execution mode.
At the infrastructure level, GraphWorkflow gained topology serialization and mid-run checkpointing for improved resilience, while AgentRearrange was upgraded for concurrent batch execution with proper task isolation. A series of critical issues were resolved as well, including YAML deserialization, shell injection, SSRF, auth cache permission flaws, silent exception handling, a max_loops iteration bug, and an Anthropic API compatibility break, resulting in a more stable and secure system overall.
Learn more about the latest features, improvements, and bug fixes in swarms v11 below ⬇️
Upgrade
Upgrade your swarms version with this command:
Github: https://github.com/kyegomez/swarms
AdvisorSwarm
AdvisorSwarm introduces a dual-agent architecture where a lightweight executor performs continuous actions while a higher-capacity advisor provides intermittent strategic guidance.
This design is inspired by Anthropic’s advisory system in Claude Code, separating execution from higher-level reasoning to improve scalability and control.
The executor runs continuously, while the advisor is invoked selectively in a budgeted manner to adjust strategy or correct course. Both share a common conversation memory, are fully provider-agnostic, and the system supports a zero-advisor mode for pure execution without oversight.
Docs: https://docs.swarms.world/en/latest/swarms/structs/advisor_swarm/
PlannerGeneratorEvaluator Closed-Loop Multi-Agent Refinement System
A structured three-agent iterative system designed to formalize generation, evaluation, and refinement into a single feedback-driven loop. It decomposes complex tasks, generates solutions, and continuously critiques outputs to improve quality over time.
This design creates a self-correcting refinement loop that converges toward higher-quality outputs across iterations, while also drawing inspiration from Anthropic’s recent work on “Harness design for long-running application development.”
Docs: https://docs.swarms.world/en/latest/swarms/structs/planner_generator_evaluator/
Grok 4.20 Heavy Mode Implementation (4-Agent & 16-Agent Variants)
We have implemented both versions of Grok 4.20 Heavy Mode, supporting a lightweight 4-agent configuration and a full-scale 16-agent cognitive mesh. Each version includes fully defined agent personalities designed for structured decomposition of reasoning, analysis, and creative synthesis tasks.
This architecture enables scalable multi-agent cognition, ranging from compact orchestration to massively parallel reasoning across specialized domains.
Swarms CLI Initialization System
A new developer-first bootstrapping flow that makes it significantly faster to spin up production-ready Swarms projects. Instead of manually wiring agents, configs, and execution entrypoints, the CLI now generates a complete runnable structure in a single command.
This change formalizes Swarms as a first-class application runtime, not just a library for composing agents.
Full docs: https://docs.swarms.world/en/latest/swarms/cli/cli_reference/
Drift Detection in Sequential Workflow
Within Sequential Workflows, you can now enable drift_detection=True to have a judge agent score the final output’s semantic alignment with the original task after each run. If the score falls below drift_threshold, the entire pipeline automatically reruns. This cycle repeats until the threshold is met.
docs: https://docs.swarms.world/en/latest/swarms/structs/sequential_workflow/
Security Hardening & Production Safety Improvements
This release significantly strengthens Swarms’ security posture across execution, file handling, networking, and CLI operations. The focus is on eliminating unsafe defaults, preventing injection attacks, and enforcing stricter isolation boundaries across all runtime utilities.
These changes collectively transform Swarms into a more production-safe multi-agent system with hardened input validation, safer system calls, and explicit network and filesystem constraints.
YAML Deserialization Hardening: Replaced unsafe yaml.load with yaml.safe_load to prevent arbitrary object injection during configuration parsing.
Shell Injection Prevention: Replaced os.system calls with subprocess.run([...], check=True) to eliminate shell injection vulnerabilities in CLI execution paths.
Bash Injection Defense Expansion: Extended command blocklists with regex-based detection to catch obfuscated payloads such as encoded execution, command substitution chains, and sensitive filesystem writes.
SSRF Protection for URL Fetching: Added strict validation to block private, loopback, and reserved IP ranges before allowing HTTP/HTTPS requests in media and image loading utilities.
Secure Auth Cache Storage: Moved authentication cache to ~/.cache/swarms/ and enforced 0600 file permissions to prevent unauthorized access.
Improved Exception Handling: Replaced silent except: pass patterns with structured logging to improve observability and debugging while preventing hidden failures.
Improvements:
AgentRearrange: O(1) Agent Lookup: Agent lookup inside `HierarchicalSwarm` was a linear scan over the agent list on every lookup. It's now a dict lookup built once at initialization O(1) regardless of swarm size.
Dependency Cleanup: Removed 4 unused packages from requirements, pinned litellm to match pyproject and more. The dependency surface is now smaller and the lockfile more stable.
Broken Documentation Links: Fixed 6 broken links in the docs and README: HierarchicalSwarm and ConcurrentWorkflow README links, 3 broken footer links in `mkdocs.yml`, and phantom entries in `structs/overview.md` for structs that don't exist (TaskQueueSwarm, MatrixSwarm, Deep Research Swarm).
Anthropic Temperature Compatibility: Removed `dynamic_temperature_enabled` from the agent parameter set. This field was being forwarded to Anthropic's API in the auto_generate_swarm.py file, which doesn't accept it, causing hard failures when using Anthropic models. The field is gone; temperature is now handled correctly for all providers.
Swarms CLI Test Suite
AgentRearrange Batch Execution: `AgentRearrange.batch_run()` now executes tasks concurrently using `ThreadPoolExecutor` with proper task isolation. Each task gets its own deep-copied agent instances to prevent state bleed between concurrent runs.
GraphWorkflow persistence introduces checkpointing, which saves workflow state to disk after each loop iteration so interrupted runs can be resumed from the last checkpoint using resume_from_checkpoint(), and serialization, which allows the full topology (nodes, edges, and agent configuration) to be exported as JSON via save_spec() and restored with GraphWorkflow.from_spec().
GraphWorkflow `max_loops` Was Always 1: A misplaced `return` statement inside the `while` loop in `GraphWorkflow.run()` caused the workflow to exit after the first iteration regardless of the `max_loops` value. The `return` is now outside the loop; per-loop results accumulate and end-point outputs are fed as context into subsequent iterations.
HierarchicalSwarm Agent Scan Improvement: Fixed agent lookup performance in `HierarchicalSwarm` O(n) scan replaced with O(1) dict lookup (same pattern as AgentRearrange fix in §5.2).
Deprecations and Removals
The following structs and utilities have been deleted. They had no active users in the codebase, their dependencies were either removed or unmaintained, and keeping them imposed ongoing maintenance cost.
SwarmTemplates: unused struct with no callers
BoardOfDirectors: superseded by HierarchicalSwarm
EuroSwarm (parliament simulation): example-only, not a production struct
MAKER: unused struct
Agent GRPO: experimental, no adoption
OpenAIAssistant / wrapper: OpenAI Assistants API deprecated upstream
check-tokens utility: dead code
data-to-text utility: dead code; callers migrated to stdlib csv
run_agents_concurrently_uvloop: no production callers, superseded by run_agents_concurrently
run_agents_with_tasks_uvloop: no production callers, superseded by run_agents_with_different_tasks
Trivy CI/CD vulnerability scanning workflow: removed from CI pipeline
If you were importing any of these directly, update your imports. The stdlib csv module is a drop-in replacement for data-to-text in all known usages.
Conclusion
This release shipped 3 new multi-agent structs, 26 improvements and refactors, and 13 bug fixes across 120 commits. The new orchestration patterns give developers more tools for decomposing hard problems across agents if that means a cost-aware advisor pairing, parallel task queues, or iterative quality loops. Alongside the new capabilities, existing structs got faster, more reliable, and more secure, with five vulnerabilities patched and over 13,000 lines of dead code removed.
Every one of these changes moves toward a single goal: building the best multi-agent framework in the world. That means not just adding capabilities, but making the existing ones more reliable, more composable, and safer to run in production. There is more to come. Stay Tuned.
Read the full changelog here: https://github.com/kyegomez/swarms/blob/master/examples/guides/v11_update/v11_changelog.md
Get started building, scaling, and deploying multi-agent workflows with Swarms:
Github: https://github.com/kyegomez/swarms
Docs: https://docs.swarms.world/en/latest/
Quickstart: https://docs.swarms.world/en/latest/quickstart/
The Swarms team and community contributors completed 120 commits in five weeks, releasing version v11, which added 3 new community architectures, 7 bug fixes, 5 security patches, removed 13,000 lines of obsolete code, upgraded multiple features, and enhanced security.
đź”— https://twitter.com/swarms_corp/status/2045975186403479887