TOOL
Build or refine a callable action with a clear input and output.
A POSSIBLE IMPROVEMENT
A repeated data-cleaning operation becomes a reusable tool.
Explore the module →GLOBAL EVOLUTION · OPEN AGENT PLATFORM
Build with agents.
Evolve the
whole system.
Plan, execute and inspect complex work in one platform. Turn discoveries from real tasks into reusable improvements across eight kinds of agent capability.
Python · Single & multi-agent · Open source / MIT
AgentEvolverHOW IT CONNECTS
YOUR TASK
An idea. A problem. A goal.
8
Evolve the capability system
One lifecycle. Eight entity types.
ToolSkillAgentWorkflowMemoryEnvironmentConnectorPlugin
Evaluate → Keep or roll back → Use again
01 / GLOBAL EVOLUTION
Eight entities.
A whole system to improve.
Choose what needs to change around the task: an action, a method, a specialist, a workflow, or the way the agent connects, observes and remembers.
SKILL
Preserve domain methods, investigation steps and useful checks as a skill.
A POSSIBLE IMPROVEMENT
A working diagnosis becomes a guide for the next investigation.
Explore the module →AGENT
Improve an agent’s instructions, capability configuration and approach. Prompts evolve with the agent.
A POSSIBLE IMPROVEMENT
Refine a reviewer’s checks and the capabilities it can use.
Explore the module →WORKFLOW
Turn a useful sequence of actions into a versioned, executable workflow.
A POSSIBLE IMPROVEMENT
Compose retrieval, parallel analysis and verification into one process.
Explore the module →MEMORY
Evolve how useful knowledge is organized and retrieved for the agent.
ENVIRONMENT
Develop stateful environments and their actions, observations and lifecycle.
CONNECTOR
Create or improve the adapters that connect an agent to external services.
A POSSIBLE IMPROVEMENT
Adapt an API response into a useful, consistent interface.
Explore the module →PLUGIN
Package integrations with configuration and lifecycle management.
Each task chooses its own scope. An improvement can involve one entity or several; prompts belong to Agent.
02 / BUILT FOR COMPLEX WORK
The foundation for
work that keeps moving.
Runtime coordinates execution. Plan preserves direction. Context carries the information needed for the next decision.
AGENT RUNTIME
Agents that collaborate.
Execution you can control.
Run focused workers, resident services and event subscribers. Exchange messages, coordinate shared resources, and pause or resume at defined execution boundaries.
Inside the Runtime →CoordinatorRuntime
WorkerBounded task
ServiceOngoing dialogue
SubscriberEvent-driven work
Messages · Resource coordination · Pause / resume · Cleanup
LIVING PLAN
Keep the next step in sight.
Maintain the goal, progress, blockers and evidence in a living plan. Its latest summary returns to the agent’s context, including after conversation compaction.
Goal & acceptance checks
Progress, blockers & next action
Detailed records, linked from a small index
Automatic planning, with an optional approval gate.
CONTEXT & MEMORY
Carry the context that matters.
Separate stable instructions, compacted history, recent turns and live state. Read memory details on demand and retain source records for inspection.
fixedInstructions & taskcheckpointCompacted historyrecentRecent interactionslivePlan, observations & budget03 / EVIDENCE-BACKED EVOLUTION
An improvement
you can inspect.
Follow a candidate from its motivation to its evaluation and later use. Versions and real call references make the decision inspectable.
- 01
Observe
Identify a need; record a baseline.
- 02
Develop
Create or refine a versioned component.
- 03
Evaluate
Compare outcomes with real calls.
- 04
Decide
Keep, roll back or unload.
- 05
Use again
Record the outcome in subsequent work.
Evaluation records link a judgment to a version and executed calls. Tasks that require verified improvement also check subsequent use.
A registered candidate is not yet proof of better results.
04 / SHARED WORKSPACE
Work alongside
your agents.
Inspect the plan, follow execution, open the files and take the next step yourself. Chat, Canvas, Code and Science share project resources.
Current workbench · A sample workspace, with no experiment running.
Pick up where the agent left off.
Science and JupyterLab use the same project kernel as the agent’s code interpreter.
Inspect what actually happened.
Read task activity, tool results and candidate details alongside project files.
Open the result. Compare versions.
Preview supported web and game deployments, with retained source versions.
05 / PUT IT TO WORK
One platform.
Different kinds of work.
Bring a bounded task and concrete acceptance checks. Domain skills shape the method; the platform supplies the common execution foundation.
WEBSITE
Build. Browse. Refine.
Implement a site, walk through it in a browser, and iterate on real interactions.
Explore →GAME
Design. Play. Iterate.
Build in Godot, inspect rendered scenes and test the experience through engine controls.
Explore →RESEARCH
Investigate. Compute. Verify.
Explore data with code, inspect outputs and continue the analysis in a shared notebook environment.
Explore →06 / ENGINEERING DETAILS
The details that
support the whole loop.
Explore how work is composed, controlled, recorded and made available for later research.
Programmatic tool use
Code Mode expresses batches, loops and branches in code, returning selected results.
Dynamic workflows
Compose parallel work, branches, loops, verification and checkpoints.
Shared budgets & permissions
Track usage across parent and child agents, with inherited permission ceilings.
Trace & recovery
Retain request and action evidence; reconcile uncertain effects before continuing.
Capabilities on demand
Search the mounted catalog and expose relevant capabilities for the next step.
Data for SFT / RL
Export execution records with provenance and reward labels for downstream training.
07 / GET STARTED
Bring a task.
Choose your capabilities.
Install the framework, configure your model provider, then launch a task. The example commands are for your local checkout.
Install the framework
git clone https://github.com/DVampire/AgentEvolver.git cd AgentEvolver bash scripts/install.sh conda activate agentosConfigure a model route
Set
LLM_HUB_API_BASEandLLM_HUB_API_KEYin your local.envfor these examples. Other providers are supported through configuration.
RUN A TASK
python examples/run_meta_agent.py \
--task "Build a useful tool and verify it." \
--cfg-options model_name=llm_hub/gpt-6-astraA configurable orchestrator with specialist agents. Review the selected config and permissions before a long run.
python examples/run_website_evolution_demo.py \
--scenario-dir examples/tasks/website_evolution/commonspace_forum \
--model llm_hub/gpt-6-astraRequires the browser dependencies and deployment setup described in the website demo guide.
python examples/run_game_development_demo.py \
--task-dir examples/tasks/game_development/tidebound_echoes \
--model llm_hub/gpt-6-astraBuild the Godot Docker images first. The default milestone is a vertical slice; --milestone campaign requests the campaign. See the game setup guide.
KEEP EXPLORING
A clear path into the code.
Start with a guide, then follow the module that owns the behavior.
A research and engineering framework for tasks worth iterating on. Start with a bounded experiment and concrete acceptance criteria.