What is the agent allowed to do?
Labs / AI as the newest chapter
The same questions now move into AI.
Labs are not startup pitches. They are where I test product beliefs under current AI pressure: agency, evidence, boundaries, memory, judgment, and human control.
Some parts are built. Some are current thinking. Some are deliberately open. The maturity boundary is part of the design.
Lab thesis
AI should make judgment more visible, not more magical.
Habitat actively tests governed autonomy. KIREI and JAVIS are earlier explorations: evidence-based product intelligence and job-intelligence directions, not validated products.
Shared questions
The labs differ, but the operating questions repeat.
What evidence is the system using?
Where does uncertainty stay visible?
Where does human judgment remain final?
Current lab directions
Three labs, shown at their actual maturity.
Each one is organized around problem, context, artifact direction, current evidence, open questions, and future direction.
Active experiment / personal AI environment
Habitat
Built and running as a personal operating environment; still evolving as a product concept.
Belief: Autonomy is not the goal. Governed help is.
01Human authority
02Permission gates
03Bounded agent work
04Observable logs
05Recovery paths
Hard approval rails External send / publishSecrets or private contextCross-profile memoryIrreversible file change
Operating loop ProposeApproveAct within scopeLog evidenceRecover or escalate
The artifact shows Habitat as an operating environment, not an AI demo: action is bounded by consent, observability, and human override.
Most AI tooling gives the assistant more reach without making the boundaries, memory, approvals, logs, or recovery paths equally visible.
If agents can touch private context, files, messages, tools, and decisions, the design problem becomes governance. Help is only useful when the human can see, steer, stop, and repair it.
The useful unit is not one powerful chatbot. It is a habitat: tools, agents, memory, routines, safety gates, model routing, logs, backups, and human authority working as one environment.
- Governed-autonomy artifact
- Agent permission map
- Memory / skill / tool boundary model
- Human override and recovery loop
- Hermes-based local agent environment
- Persistent memory and skill system
- Scheduled routines and watchdogs
- Discord operating channel
- Local-first private context indexes
- Approval and safety-gate habits
- Which actions should require hard approval every time?
- How should agents explain uncertainty without creating noise?
- What should be local-only by default?
Turn the current personal operating environment into a clearer external-facing system model: bounded autonomy, observable work, reversible action, and human authority as the center.
Active product exploration / evidence-based product intelligence
KIREI
Active product exploration. It is not presented as validated, adopted, or launch-ready.
Belief: Trust improves when evidence helps people understand tradeoffs, not when a product pretends certainty.
01Product claim
02Ingredient context
03Evidence confidence
04Personal fit
05Decision support
Beauty, wellness, and household products are hard to judge because ingredient lists, claims, local availability, personal context, and cultural norms are disconnected.
A blunt safe/unsafe score can create false certainty. People need evidence that helps them decide what matters for them: sensitivity, context, tradeoffs, source credibility, and local alternatives.
KIREI should behave less like a judge and more like an evidence layer: show what is known, what is uncertain, why it may matter, and how the answer changes by person or use case.
- Evidence confidence card
- Ingredient-to-context map
- Personal fit decision frame
- Concept domain framing
- Evidence-first product principles
- Ingredient-context model direction
- Personal-fit decision-frame direction
- What sources are credible enough for consumer guidance?
- How should uncertainty be displayed without alarming people?
- How can JP/Asia context remain specific instead of generic?
Continue shaping a narrow product lookup prototype around evidence, uncertainty, personal-fit prompts, and safer language before making any validation or launch-readiness claims.
Active product exploration / job intelligence
JAVIS
Active product exploration. It is not presented as a validated product, traction story, or production platform.
Belief: Career AI should improve judgment and agency, not maximize application volume.
01Market signal
02Role evidence
03Fit hypothesis
04Effort decision
05Human choice
Job search tools optimize listings and applications, but people still struggle to understand role fit, market signals, company context, timing, and where effort is worth spending.
More applications can make candidates busier without making them wiser. A useful job-intelligence system should improve judgment: what to pursue, why it fits, what evidence is missing, and when to stop.
JAVIS should connect hiring-system knowledge with job-intelligence workflows: role discovery, signal tracking, application strategy, company context, and reflection without turning the person into an application machine.
- Role-fit evidence board
- Opportunity signal map
- Application decision gate
- Job-intelligence concept framing
- Role-discovery model direction
- Market-signal monitoring direction
- Connection to hiring workflow expertise from Indeed/SEEK work
- What makes a role worth pursuing beyond keyword match?
- How much should an agent automate versus prepare?
- How should rejection/no-response data inform strategy without damaging confidence?
Shape a small job-intelligence workflow that tracks selected signals, explains fit, recommends effort level, and keeps the final decision human before making any production-platform claims.
Guardrail
AI is the newest chapter, not the origin story.
The labs are strongest when they point back to the professional work: Virtual Recruiter’s trust/delegation model, EMA’s signal-to-direction work, Flash Funnel’s decision-load reduction, and Indeed University’s leadership system.