Skip to content

Latest commit

 

History

History
63 lines (47 loc) · 2.88 KB

File metadata and controls

63 lines (47 loc) · 2.88 KB

How steward compares

vs OpenClaw / persistent autonomous agents

OpenClaw and similar systems emphasize persistent autonomous iteration: the agent runs continuously, deciding what to do, executing, and looping.

Charter-worker takes a different approach:

OpenClaw-style Charter-worker
Execution Continuous autonomous loop Scheduled cycles (hourly/daily/weekly)
Control Agent decides everything Charters, guardrails, human feedback loops
Cost model Constant high LLM usage Amortized: routine work runs as cheap scripts
Recovery Agent self-corrects (or doesn't) Orchestrator diagnoses, fixes, retries
Observability Logs Structured summaries, email reports, status surface

The key difference is amortized autonomy: repeated work should crystallize into cheap explicit workflows, not stay in permanent high-cost agent mode.

vs autoresearch / FARS / ARIS / Elicit

These are research-specific pipelines with hard-separated stages: search → read → summarize → (maybe) write.

Charter-worker differs in several ways:

Research pipelines Charter-worker
Scope Research only Any recurring task
Pipeline Fixed stages (search → read → write) Interleaved: search, act, experiment, escalate
Experiments Not supported Built-in experiment dispatch + result parsing
Feedback One-shot or manual iteration Email feedback loop, reply-to-steer
Self-healing None Diagnostic agent on crash
Long-horizon Single session State persists across cycles with 10 guardrails
Theory Ad-hoc pipeline design Grounded in organizational theory (Coase, Simon, Galbraith)

vs general task runners (cron, Airflow, Prefect)

Charter-worker is not a workflow orchestrator in the traditional sense. It does not have DAGs, task dependencies, or a web UI.

What it adds over plain cron:

  • Agent dispatch: spawns LLM agents (codex, claude) as subprocesses
  • Lock management: prevents double-spawning across hourly cron cycles
  • Preflight checks: validates prerequisites before running
  • Self-healing: diagnoses and fixes crashes automatically
  • Structured summaries: every task produces summary.json + summary.md
  • Email digest: collects all results into a single daily email
  • Workflow promotion: tasks can graduate from expensive agent mode to cheap scripts

What steward is NOT

  • Not a web product or SaaS
  • Not a replacement for Airflow/Prefect for production data pipelines
  • Not a general-purpose AI assistant (it runs your recurring tasks)
  • Not a multi-agent framework (it orchestrates tasks, not agent conversations)

What it IS

A governed autonomy engine for recurring tasks, grounded in organizational theory, designed for a single researcher or developer who wants their AI agent to handle routine work reliably and cheaply.