Cognition — the live loop
A continuous control loop that senses, reasons, and acts with guardrails. Every iteration evaluates goals, constraints, uncertainty, and cost.
goal stackrisk modelbounded autonomy
Core Reasoning Loop
Policy Engine
Session Memory
AGS++ Policies
Reasoning — plans, proofs, and checks
Multi-pass reasoning blends structured planning and deterministic checks. Plans are validated against policy before execution.
validationcritiqueself-checks
RAG++ — retrieval that plans
RAG++ builds a retrieval plan: sources, filters, joins, ranking. Each citation carries provenance and freshness.
provenancerankingtemporal scope
LLM Synth — structure, not just text
LLMs draft structured artifacts—plans, tables, JSON, code—validated by schemas. Invalid outputs are repaired automatically.
schemarepairdeterminism
Execution — tools with policy
Actions are transactional: typed inputs, dry-runs, approvals when required. Every side effect is logged with policy context.
dry-runapprovalstrace
Perception — files, vision, audio, sensors
Unified inputs across documents, images, audio, and sensors. Each read is scoped by policy and redacted if out of bounds.
multimodalredactionPII-aware
Memory — session, episodic, and graph
Short-term context meets durable knowledge. Episodic memory is searchable; graph captures entities and relationships.
contextvector+graphTTL/retention
Kernel — scheduling and isolation
Workloads run as spans with quotas and priorities. Sensors and tools are isolated with backpressure and retries.
quotasprioritybackpressure
DNA — roles, templates, inheritance
Org DNA encodes roles, capabilities, and defaults. Policies inherit from DNA so new agents start safe and consistent.
rolescapabilitiesinheritance
Platform
Cognitive OS with shared context
A live reasoning kernel orchestrates agents over the Signal Bus with backpressure, transactional tool calls, and signed decision traces. Run in cloud, on-prem, or edge.
- Streaming context fabric with drift detection
- Typed tool interfaces and dry-run mode
- Observability: spans, risk, and outcomes
Agents
Specialists that think and act
Compose domain experts, wire their tools, and let the Cognitive Loop coordinate decisions.
Ops & Routines
Schedulers, service timetables, incident triage, runbook executors.
ApprovalsSLA-aware
Vision & Robotics
Thermal/vision detectors, patrols, manipulation, Orin/Edge control with guardrails.
ROSEdge
Data & Knowledge
ETL, schema induction, graph synthesis, retrieval planning, report drafting.
ProvenanceRAG++
Compliance
Policy-first safety (AGS++)
Policies govern what agents may see and do. Local data stays local; guardrails apply per policy and inherited policy to guarantee bounded outcomes.
- Allow/deny lists, scoped limits, approvals
- Local processing with hierarchical policies
- Signed audit trail for every step
{
"role": "cleaning-supervisor",
"allow": ["read:occupancy", "create:work_order"],
"deny": ["delete:work_order"],
"requireApproval": [{"action":"dispatch_worker","by":"supervisor"}],
"limits": {"zoneRadiusM":50, "priorityMax":3}
}FAQ
Common questions
What does “bounded autonomy” mean?
Agents can act independently inside defined policies: limits, approvals, and scopes. Attempts outside bounds are blocked or escalated.
How does Pegesis keep decisions explainable?
Every step (retrieval, reasoning, tool call) is traced with inputs, outputs, policy context, and citations—exportable for audit.
Can AGS++ run at the edge?
Yes. The kernel supports quotas, offline operation, and backpressure; policies travel with the workload.