Console

The ARISE Console is a web UI for creating agents, watching them evolve tools in real-time, and inspecting evolved code.

arise console
# Opens http://localhost:8080

Features

Agent Management

Create and manage multiple agents from the browser. Each agent has its own skill library, trajectory store, and configuration.

• Create agents — pick a model (OpenAI, Anthropic, Bedrock), write a system prompt, choose a reward function
• Per-agent stats — skills count, success rate, episode count, evolution count
• Live status — running, stopped, evolving

Live Terminal Feed

Watch your agent work in real-time. The Live tab shows a terminal-style feed of episodes and evolution events streamed over WebSocket.

• Episodes with color-coded status (green OK, red FAIL)
• Evolution blocks: gap detection, synthesis progress, test results, promotion
• Run tasks interactively from the input bar
• Events persist across page reloads

Skill Inspector

Click any evolved skill to see:

• Full implementation with syntax-highlighted Python code
• Test suite with standard and adversarial test results
• Performance metrics: success rate, invocations, latency, version
• Export as .py or deprecate

Editable Configuration

Change agent settings on the fly from the Config tab:

• System prompt
• Reward function (task_success, answer_match, code_execution, efficiency, LLM judge)
• Failure threshold
• Sandbox backend

Changes take effect on the next run — the ARISE instance is recreated with the new config.

Global Views

• All Skills — table of every skill across all agents, sortable by success rate
• Evolution Log — timeline of evolution cycles with gaps detected, tools promoted/rejected

Configuration

# Custom port
arise console --port 3000

# Custom data directory
arise console --data-dir ./my-arise-data

# Custom host
arise console --host 127.0.0.1

Agent data is stored in ~/.arise/console/ by default:

• agents.json — agent configs (persisted across restarts)
• agents/<id>/skills/ — SQLite skill library per agent
• agents/<id>/trajectories/ — SQLite trajectory store per agent
• agents/<id>/events.jsonl — live feed event log

Reward Functions

The Console includes these reward function presets:

Function	Description
task_success	Checks metadata signals (expected output, success flag). Falls back to 1.0 if no errors.
answer_match_reward	Compares output against expected answer. 1.0 exact, 0.7 substring, 0.0 no match.
code_execution_reward	1.0 if no tool errors, minus 0.25 per error.
efficiency_reward	Penalizes extra steps. 1.0 for 1 step, -0.1 per additional.
llm_judge_reward	LLM rates response quality 0-1. Uses the agent’s own model. ~$0.001/call.

Model Support

The Console auto-maps short model names to provider-specific IDs when AWS credentials are available:

Short Name	Maps To
claude-sonnet-4-5	bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0
claude-sonnet-4	bedrock/us.anthropic.claude-sonnet-4-20250514-v1:0
claude-haiku-4-5	bedrock/us.anthropic.claude-haiku-4-5-20251001-v1:0
gpt-4o	gpt-4o (requires OPENAI_API_KEY)
gpt-4o-mini	gpt-4o-mini (requires OPENAI_API_KEY)

You can also use full provider-prefixed names directly: bedrock/..., anthropic/..., openai/....