GrokBot supervises
Intent, preferences, tradeoffs, context, and final decisions stay in a small persistent control layer.
Open-source workforce architecture / v0.1.0
A low-usage workforce architecture for turning GrokBot into a persistent control plane instead of an expensive universal worker.
Public source: M4G3LL4N0/grokbot-office · AgentOS: M4G3LL4N0/agentos
The thesis
GrokBot supervises
Intent, preferences, tradeoffs, context, and final decisions stay in a small persistent control layer.
AgentOS routes
Capabilities, workers, retries, verification, and learning are selected at the execution boundary.
Workers execute
OpenCode, models, APIs, MCP, browser workers, and deterministic software remain replaceable.
Evidence verifies
The system learns and reuses only after a result crosses a real evidence boundary.
Interactive architecture simulation
Selected route / SOURCE CHANGE
request
cache state
changed files
OpenCode
tests
verifier
cache
result
Illustrative result
Compact audit + evidence
The second request can retrieve the verified artifact when the source hash is unchanged.
Architecture simulation / not live execution evidence
North star
Verified useful output
÷
Total resource cost
The point is not to make every task expensive. It is to make the next equivalent task cheaper by retrieving the right artifact, reducing context, and keeping the smallest capable team in control.
134
conceptual roles
3
reference supervisors
0
fabricated benchmarks
1
smallest-capable rule
The learning loop
A worker does not get to call its own output “verified.” The loop keeps the boundary visible.
Selected: Worker
Start with the smallest capable owner, not a fan-out.
Role catalog / sanitized reference data
Own attention, merge chiefs, and surface decisions.
Own external intelligence, changes, anomalies, and opportunities.
Know project states, blockers, dependencies, and next actions.
Coordinate personal logistics and planning with approval boundaries.
Coordinate specialist workflows for assets, admin, and risk.
Turn questions into minimum-cost research plans.
Coordinate repeatable software and operational automations.
Find lower-cost paths without weakening safety or evidence.
Challenge outputs and require evidence before reuse.
Protect identities, scopes, and approval boundaries.
Retrieve approved place context and calculate useful options.
Discover and inspect external capability candidates without executing them.
Read the operating model