Planner
Defines bounded steps instead of allowing an unstructured request to become arbitrary execution.
AGENT SYSTEMS / GOVERNED AUTOMATION
Project MIND Agent Workbench
A local-first, deterministic agent workflow that makes planning, evidence, verification, and proposed actions inspectable.
This original Joshua Randall implementation demonstrates the orchestration and safety plane proposed for Project MIND. It turns a technical goal into bounded tasks, records evidence requirements, checks whether claims are supported, and produces a dry-run action list.
Verifier
Blocks completion claims when evidence is missing and records uncertainty in the ledger.
Executor
Produces a proposed action plan and requires explicit human approval before side effects.
The workbench is intentionally deterministic and local. It does not claim consciousness, general intelligence, autonomous self-improvement, or authority to modify external systems.
Problem, implementation and evidence
Evidence-looking words cannot independently verify a claimed outcome.
The existing workflow preview organizes a dry-run plan; the new companion compares actual typed artifacts with explicit acceptance criteria and tests replay.
What the evidence establishes
Deterministic local workflow and executed in-memory coordination fixtures; no live LLM agents, external tools or distributed consensus run.
Reproduce or extend
Add persisted attempts, artifact schema versions and stale-result rejection before connecting a new external execution boundary.
Open the related experiment or review surface →