BUILD

Agents

Autonomous task executors within your Spaces.

How agents work

An agent receives a task, decomposes it into steps, and executes each step using the available Skills. Agents read from and write to the active Space — so every file they touch, every memory they store, and every embedding they create is scoped to your project.

TARX Local

TARX Local includes a built-in agent executor. It watches for pending tasks and spawns Claude Code instances to handle them.

# Start the agent executor
tarx daemon start

# Submit a task
tarx daemon task "Write unit tests for auth.ts"

# Check status
tarx daemon status

Safety

Max 2 concurrent tasks
5-minute timeout per task
3 retries before auto-pause
Auto-pause on 3 consecutive failures
Full audit log at ~/Library/Application Support/tarx/audit.jsonl

Research agent

tarx_research is a specialized agent that runs autonomous experiment loops. It modifies a target file, runs a timed experiment, evaluates a metric, and keeps or discards the change. Repeats until budget is exhausted or the metric converges.

tarx research \
  --experiment-file train.py \
  --program-file sweep.md \
  --metric val_loss \
  --budget 300