How it works
A reasoning CEO that decides how to think before it answers
The CEO agent plans, stress-tests its own plan, decides whether to work alone or assemble a team, executes, reflects, synthesizes, and passes a final safety gate — every time.
The 6-node decision loop
Plan
Decompose the task and decide serial vs. swarm.
Meta-critique
Stress-test the plan before acting.
Execute
Run directly or coordinate specialists.
Reflect
Loop back on low confidence.
Synthesize
Assemble a single governed deliverable.
Safety gate
Final governance + guardrail check.
Decomposition
Serial when simple, a swarm when it isn't
The CEO evaluates each task and chooses the cheapest path that still meets quality — answering directly, or decomposing the work across specialist agents it coordinates.
Serial mode
For tasks the CEO can handle directly, it runs the loop itself — no specialists spawned, minimal overhead.
Swarm mode
For complex work, the CEO decomposes the task, spawns specialists (each inheriting curated knowledge), coordinates them, and reconciles their outputs into one deliverable.
Orchestration is pluggable: LangGraph is the default backbone, with Swarm and Hybrid strategies available, and new strategies added by registering a subclass — no changes to the CEO's core code.
Specialists
Specialists that inherit what they need
When the CEO spawns a specialist, it passes down four pillars of context so the agent starts informed — not from a blank slate — within a managed VRAM budget.
Inference or container
Specialists run in inference mode or as containers, depending on the work.
4-pillar inheritance
Curated knowledge, contracts, and context are handed down at spawn time.
VRAM budgeting
Spawning respects a managed memory budget across the GPU fleet.
The research panel
Evidence specialist, analysts, composer, evaluator, chart planner, and renderers.
Inference
A dual inference path
Local NVIDIA NIM serving powers the CEO and most specialists; an API tier with a multi-provider fallback chain handles the research roles that need very large context windows.
Local NIM tier
Per-role precision routing (FP8 default), TensorRT-LLM-aligned controls, and a VRAM-aware downgrade path keep local inference fast and within budget.
Large-context API tier
Research roles that need very large context windows use API LLMs, with a circuit breaker, retries, and a multi-provider fallback chain.
Governance
The safety gate is part of the loop
Risky actions escalate for human review and oversight relaxes only as the system earns trust — a graduated-autonomy model the program calls "Iron Dome".
Baseline
Maximum oversight; risky actions escalate for human review.
Learned trust
Oversight relaxes as the system demonstrates reliability.
Mature
Earned autonomy within rubric-based, auditable limits.