Autonomy needs a boundary
Agents receive explicit roles, permissions, escalation paths, and stop conditions before they receive tools.
Human-governed autonomy for serious operators
Let your AI agents run.
Keep judgment in command.
TonyCAPM designs closed-loop AI operating systems that coordinate research, execution, verification, and stop conditions—without turning autonomy into uncontrolled risk.
18yearssemiconductor systems
10+yearsquant research
8,100+documents indexed
HHuman
judgmentauthority layer
judgmentauthority layer
01Observe
02Route
03Execute
04Verify
05Stop
AIOS · CLOSED LOOP / OPEN EVIDENCE
01 / THE OPERATING THEOREM
Capability is accelerating. Accountability is not.
Frontier models may outperform the harness. They still cannot inherit the consequences.
Harnesses, agents, and models will keep improving. TonyCAPM does not claim to slow that frontier or own it. The enduring work is to define the operating constitution around it: what the system may observe, when it may act, how its work is verified, and where a human must stop it.
Memory needs provenance
Research and decisions become shared evidence—traceable across agents, devices, and time.
Execution needs verification
Outputs do not become actions merely because a model produced them. Tests and risk gates close the loop.
Humans retain consequence
The operator owns intent, material risk, and the irreversible call. AI expands judgment; it does not absorb liability.
02 / THE WORKING SYSTEM
AIOS-LITE · EDGE CONTROL PLANE
A multiplayer operation that can keep moving when the founder steps away.
The architecture separates continuity from compute. A lightweight, always-on control node preserves intent, safety, and shared state. Heavier machines handle models, databases, simulations, and bulk research.
Human layerPurpose · approval · kill switch
CONSTITUTIONSOUL.md
Values, boundaries, and decision posture.
CONTROLOpenClaw
Routing, safety gates, and operator interface.
Continuity layerAlways on · shared context
QUEUEAgent Board
Work state and handoffs.
LEDGERTS-Bridge
Cross-agent decisions in Git.
MEMORYPensieve
Durable research context.
RECOVERYOps LTSSM
Health, drift, and return.
Execution layerElastic · replaceable compute
RESEARCH & VALIDATIONPython / R runners
Backtests, simulations, falsification, and evidence.
HEAVY ORCHESTRATIONWindows / WSL research plane
Containers, databases, dashboards, ingestion, and local models.
Loop posture: autonomous where reversible · human-gated where consequential
03 / START SMALL, LEARN FOR REAL
The first product is clarity.
An AIOS readiness session turns agent enthusiasm into an operating decision.
01 · DIAGNOSTIC
AIOS Readiness Session
A focused working session for a founder or technical operator already using multiple AI tools.
- Map your agents, models, data, and execution paths
- Identify unsafe handoffs and invisible human middleware
- Define approval gates, evidence, and a first closed loop
02 · PILOT
One Governed Loop
Implement one bounded workflow—research, engineering analysis, or operational reporting—with measurable verification.
OutcomeA working loop, not a transformation deck.
03 · OPERATING DESIGN
AIOS Control Plane
Evolve successful loops into a lightweight coordination architecture with shared memory, task state, and recovery.
OutcomeMore autonomy without surrendering legibility.
04 / FIELD NOTES
Evidence, experiments, and operating judgment.
Your operating notes and public AI research essays live at the TonyCAPM Research Blog. This homepage remains the AIOS front door; the blog remains the evidence trail.
Read the research blog ↗TC
05 / A NOTE FROM TONY
“The machines may become better than my harness. That is expected. My work is to make sure the organization becomes wiser—not merely faster.”
I am Tony C. Chen, a semiconductor systems and quality practitioner in Taipei who became a quantitative researcher and AI operating-system builder. I bring the same instinct to all three domains: make the process visible, test the uncomfortable failure mode, and keep a human accountable for what happens next.
For founders and operators building beyond the demo.