AgentEvolve is Avernet’s local workspace for evaluating and continuously improving personal OpenClaw Bots. It connects real sessions, diagnosis, Bench evaluation, optimization, and recoverable versions into one traceable workflow.
For Chinese documentation, see agent-evolve.zh-CN.md.
Historical source paths, script names, and configuration fields may still contain
clawwebfor compatibility. AgentEvolve is the public product name.
| Capability | Purpose | Open-source edition |
|---|---|---|
| Bot diagnosis | Find issues in personal-Bot sessions, extract good/bad cases, and produce a follow-up plan | Available |
| Bench diagnosis | Evaluate a Bot with published local templates and inspect cases, metrics, sessions, and raw output | Available |
| Diagnosis-driven optimization | Continue from a completed diagnosis and plan | Available |
| Bench optimization | Establish a baseline and optimize against train/test domains | Available |
| Bot self-evolution | Run diagnosis-first or direct-goal full flows | Available |
| Pack and Apply | Create a recoverable local snapshot and apply a previous version | Available |
| Task cleanup | Remove evolution agents and sessions carrying explicit task markers | Available |
| Task and version management | Inspect steps, outputs, evaluation results, diffs, and Pack versions | Available |
Service-Bot flows, dedicated session diagnosis, Bot repair, governance optimization, and internal-environment integrations are not included in the current open-source edition.

The task center brings tasks, evaluations, and versions together while exposing only the capabilities available in the current edition.
Tasks use only the stages they need. Bot diagnosis is read-only, while Bench diagnosis evaluates without running an optimization loop.
~/.openclaw containing openclaw.json and workspace/.From the Avernet repository root:
cd apps/evolverun/clawweb
npm ci
cd ../../..
bash apps/evolverun/clawweb/scripts/start-clawweb-open.sh

The interactive launcher selects the Bot source and prepares the local AgentEvolve runtime.
The launcher asks which Bot source to use:
~/.openclaw;~/.openclaw directly without starting Singlebox;Explicit examples:
# Use ~/.openclaw directly
bash apps/evolverun/clawweb/scripts/start-clawweb-open.sh \
--bot-source openclaw
# Reuse an Avernet Singlebox Bot database
bash apps/evolverun/clawweb/scripts/start-clawweb-open.sh \
--bot-source singlebox \
--user-id mock-user
# Select a model and port
bash apps/evolverun/clawweb/scripts/start-clawweb-open.sh \
--bot-source openclaw \
--model provider/model \
--port 5173
Open http://127.0.0.1:5173/ and select a personal Bot.
The launcher starts only AgentEvolve. It does not start, stop, or restart Avernet Singlebox or an OpenClaw gateway. If port 5173 is occupied, it exits instead of replacing the running process; use --port to select another port.
For the first run, avoid making changes to an important Bot; use an objective that is easy to verify and roll back.
Use this when the desired behavior is already clear.

Select the personal Bot, evolution mode, execution model, and objective when creating a task.
The planning stage converts the user’s goal into an executable Spec and Bench cases. Acceptance is based on actual validation, not literal text matching between the input and generated documents.
Use diagnosis when the root cause is unknown.
Diagnosis reads session history and must not modify the target workspace.
When report generation is not configured, AgentEvolve shows a neutral unavailable state; the evaluation data remains available.

Bench pages retain the template, version, and historical metrics for repeatable comparisons.
Applying a Pack changes the target Bot. Verify the Bot, Pack source, diff, and evaluation result first, and keep a rollback version.

Compare evaluation results, inspect diffs, and download or apply recoverable Packs from the version page.
Do not accept a version based on aggregate score alone. Critical cases and the actual diff also require review.

The task detail groups planning, optimization rounds, diffs, Bench metrics, and Packs for step-by-step review.
| Bot source | Default task directory |
|---|---|
| Avernet Singlebox | ~/.local/share/clawweb-singlebox |
| Local OpenClaw | ~/.local/share/clawweb-open |
Pack files are stored under dataDirectory/artifacts. The protocol retains an oss:// logical reference, but the public filesystem object store resolves it to a local file; it does not upload to an internal object store.
Use --data-dir to override the task directory. Different histories after switching Bot sources are expected because the sources use separate databases and data directories by default.
The launcher generates a machine-local runtime configuration. User-owned YAML/JSON configuration must stay outside the repository and must not contain committed credentials.
| Field | Required | Meaning |
|---|---|---|
userId |
yes | Local user identifier |
model |
yes | Initial model; otherwise read from the OpenClaw default model |
botSource |
no | singlebox or openclaw |
backendDb |
in singlebox mode |
SQLite database created by Avernet Backend/Singlebox |
openclawHome |
in openclaw mode |
Directory containing openclaw.json and workspace/ |
dataDirectory |
yes | Task, log, and Pack artifact storage |
skillsRoot |
no | Public Skills root; defaults to the Skills shipped in Avernet |
botsRoot |
no | Custom local Bot data root |
port |
no | Web port; default 5173 |
maxArtifactBytes |
no | Maximum streamed artifact size |
Model choices come from the AgentEvolve module configuration, and the open-source edition also accepts a custom provider/model. Credentials remain owned by the model provider or OpenClaw configuration.
The public implementation is under apps/evolverun/clawweb; public runtime Skills are under apps/evolverun/clawweb-skills/clawevolve-skills.
Public functionality must not depend on private services, endpoints, or credentials. Environment-specific values belong in arguments, configuration, or environment variables. Runtime data must not be written into the source tree. Workspace changes, Pack application, and cleanup must remain path-bounded and traceable.
Useful checks:
bash apps/evolverun/clawweb-skills/clawevolve-skills/scripts/verify_public_skills.sh
cd apps/evolverun/clawweb
npm run check
npm run build
npm test
backendDb, userId, and openclawHome; only personal Bots owned by the current user are listed.singlebox/openclaw or dataDirectory changed.verify_public_skills.sh and restart from the same Avernet checkout.node_modules and run npm ci on the current OS/CPU with a supported Node/npm version.--port.