Elastic License 2.0

Asimov

Dynamic Agentic AI
with Human-on-the-Loop

Claude provides velocity. Asimov provides guardrails.

39x better than fixed agentic at 10 steps. 1502x at 20.

38k

Lines of Code

429

Tests

Releases

Protocols

200k

Context Tokens

Human-on-the-Loop

Not Human-in-the-Loop. Not fully autonomous.
The sweet spot for AI that ships.

HITL (Human-in-the-Loop)

Human approves every decision
High latency (blocks on human)
Limited by human attention
Safe but slow

HOTL (Human-on-the-Loop)

Human monitors, intervenes when needed
Low latency (AI operates autonomously)
Scales with AI capability
Fast and safe

HOTL catches errors between spawns.
75% error detection rate vs 40% for fixed agentic.

Dynamic Swarm Architecture

One orchestrator with maximum context.
Sub-agents spawned dynamically at runtime.

Human (HOTL)
    ↓ oversight (can intervene at any step)
Orchestrator (~200K tokens, extended thinking)
    ↓ spawns dynamically at runtime
    ├── Sub-Agent 1 (~200K tokens)
    ├── Sub-Agent 2 (~200K tokens)
    └── Sub-Agent N (~200K tokens)

Each sub-agent gets full context, not fragmented 8-32K chunks.
AI decides topology at runtime. No fixed roles.

The Math

Monte Carlo simulation. 10,000 trials.
Validated against R and Gnumeric.

At 10 Steps

39x

better than fixed independent agents

At 20 Steps

1,502x

advantage ratio over fixed agentic

Steps	Dynamic Swarm + HOTL	Fixed Centralized	Fixed Independent
5	95.2%	49.9%	15.3%
10	90.7%	24.9%	2.3%
20	82.2%	6.2%	0.05%
50	61.3%	0.1%	~0%

Source: Google/MIT Research (Dec 2024) — 17.2x error amplification for independent agents, 4.4x for centralized.

8 Protocols

Hardcoded, not configurable. By design.
Guardrails that can't be prompt-injected away.

asimov

Three Laws. Harm prevention. Human veto commands: stop, halt, abort.

sycophancy

Truth over comfort. Disagree openly. No false agreement.

freshness

MUST use ref for web fetching. Never WebSearch/WebFetch (403 blocked).

sprint

Autonomous execution until completion. Don't stop to ask — document and continue.

warmup

Session initialization. Load protocols, project, roadmap. Zero file reads.

green

Efficiency awareness. Warn on less efficient language/framework choices.

coding-standards

Human-readable, beautiful code. Tests are documentation. No warnings.

migrations

Functionally equivalent transformations. For migration-type projects only.

Zero File Reads

One command. Complete context.
asimov warmup

# Session start hook outputs JSON:
{
  "project": { /* full project.yaml */ },
  "roadmap": { /* full roadmap.yaml */ },
  "protocols": { /* all 8 protocols */ },
  "wip": {
    "active": true,
    "next_milestone": "10.5.0"
  }
}

Claude knows your project, patterns, and ethics before you type anything.
No cold starts. No context loss.

vs Fixed Agentic Frameworks

Dimension	Dynamic Swarm + HOTL	LangChain / CrewAI / AutoGen
Context per agent	~200K tokens each	8-32K fragmented
Agent spawning	AI-decided at runtime	Pre-defined at design time
Human oversight	HOTL gate between steps	None or batch approval
Error amplification	Contained (HOTL catches)	17.2x (independent)
Max effective agents	Unlimited (AI-managed)	3-4 (Rule of 4)
Success at 20 steps	82.2%	0.05 - 6.2%

Source: Cognition (Devin) — "Multi-agent systems in 2025 result in fragile systems. Decision-making too dispersed."

Get Started

Install and initialize in 30 seconds.

# Install
cargo install royalbit-asimov

# Initialize project
asimov init

# Start session with full context
asimov warmup

View on GitHub

License

Code: Elastic License 2.0 — Source available, not open source

✓ Personal and internal use
✓ Commercial use (non-competing)
✓ Modify for internal use
✗ Provide as managed service
✗ Circumvent license keys

Documentation: CC BY-NC-ND 4.0

For commercial licensing, open a GitHub issue.