INSOMNIQ
initializing system...
Give INSOMNIQ a problem. It assembles a team of autonomous AI agents — Claude Sonnet, Haiku, Gemini, or local Ollama — designs the pipeline, executes in parallel, and self-improves with every run. Not a wrapper. An AI firm.
Type a problem, watch the team form
Follow on GitHubHow it works
You describe what needs to happen. INSOMNIQ assembles a specialist team, designs the pipeline, executes in parallel, evaluates the quality, and iterates — until the goal is met. Fully autonomous, end to end.
01
Mission Brief
YAML or plain language. Set the objective, budget ceiling, acceptance criteria, and any constraints. That's it — INSOMNIQ takes full ownership from the first token. No scaffolding, no hand-holding, no prompt engineering required.
02
AI Strategy
The Coordinator reads the Knowledge Base for proven strategies from past runs. It picks the agents, assigns models — Claude Sonnet for deep reasoning, Haiku for speed — and blueprints the entire pipeline before a single node fires.
03
Parallel Execution
Independent nodes execute simultaneously. The Predictor queries the KB for known failure patterns before each node starts — intercepting risks before they hit. The Auditor monitors every output in real time, reading the inbox for your signals.
04
Quality Gate
The Evaluator scores output 0–100% against your exact acceptance criteria. For complex deliverables, it spawns parallel sub-evaluators. If the score falls short, it triggers an automatic replan — with a full diagnosis of why it failed.
05
Compound Learning
The Librarian writes structured insights to the Knowledge Base after every run — what worked, what failed, which models excel at which task types. The Coordinator reads this on the next run. The system doesn't just repeat. It evolves.
The AI Agent Team
Each agent is an LLM-powered autonomous specialist. They don't take turns — they run in parallel, communicate through signals, and compound knowledge across every run via UCB1 model selection.
Planner · Orchestrator
The brain of every run
Plans the entire pipeline. Reads KB for proven strategies, designs agent teams, assigns LLM models per node, and handles replanning when something fails.
Observer · Watchdog
Never misses an anomaly
Watches every node run. Manages the inbox — users and Investor send signals here. Delivers consolidated summaries to Coordinator at each replan.
Risk-Taker · Explorer
Breaks safe-play loops
When quality stagnates, calls a board meeting with concrete goals. Reads quality signals and forces the system to take calculated risks or reboot entirely.
Judge · Quality Gate
Your quality bar, automated
Scores output quality 0–100% against your acceptance criteria. Spawns parallel sub-evaluators for complex output. Decides if can_improve — triggering the next iteration cycle.
Memory · KB Keeper
LanceDB + ModernBERT, 3ms search
Consolidates the Knowledge Base after every run. Merges duplicates, drops noise, and surfaces actionable insights for the next Coordinator.
Forecaster · Risk Scout
Future failure prevention
Before each node runs, queries KB for past failures on similar tasks. Flags risks in advance — so the pipeline doesn't repeat history.
Meta moment
Right now — attempt 19. Pipeline: web_spark → html_architect → qa_polish. Budget: $22.52 of $80.00 spent. Not a demo. Actual autonomous execution, live.
This page was created by INSOMNIQ — 20 iterations, $28, 92% quality
19
ATTEMPT
$23
SPENT OF $80
3
PIPELINE NODES
Live session
A real INSOMNIQ pipeline session. Parallel LLM execution — Claude Sonnet for strategy & code, Claude Haiku for speed & memory, Ollama for local runs. Autonomous AI agents, unfiltered.
Why INSOMNIQ
Other tools start from zero on every session. INSOMNIQ has a persistent, searchable knowledge base that grows with every run — making each iteration fundamentally smarter than the last.
The old way
Stateless execution
Every session starts from scratch. No memory of what worked, what failed, or which model performed best.
Full restart on failure
One bad node kills the whole pipeline. You restart from zero — burning cost and time.
No quality measurement
Output lands in your hands and you decide if it's good enough. No score, no criteria, no feedback loop.
Fixed model, fixed strategy
Same model for every task. No learning about which model fits research vs. coding vs. writing.
Code tasks only
Built for developers — not for research, writing, analysis, strategy, or general knowledge work.
INSOMNIQ
Persistent Knowledge Base
LanceDB + ModernBERT vector store. Every run deposits insights — strategies, failures, model performance data. 3ms semantic search.
Blame-driven selective retry
When a node fails, only that node replans. The Coordinator surgically fixes the root cause — no wasted compute, no full restart.
Automated quality scoring
Evaluator scores 0–100% against your criteria. Spawns parallel sub-evaluators for complex outputs. Quality is measurable, repeatable, automatic.
UCB1 adaptive model selection
Agent Resources learns which model performs best for each task type across runs — chosen from evidence, not defaults.
Any knowledge task
Research, analysis, writing, coding, data processing, strategic planning. If an LLM can do it, INSOMNIQ can automate it with a full agent pipeline.
672
Tests Passing
17.5K
Lines of Code
3ms
KB Search
9
Core Pillars
Community & Roadmap
INSOMNIQ isn't just another AI wrapper. The shared Knowledge Base and public patterns compound intelligence across the entire community — while your local data stays completely private.
Investor Center
Cross-project knowledge aggregation. Your Investor learns from every project in your portfolio. Public Investor Center lets the community share collective AI team patterns and strategies.
Public Knowledge Base
Opt-in shared KB. Strategies, agent patterns, domain knowledge — contributed across the community. Your local KB stays private. Public KB is read-only unless you choose to contribute.
Server Mode
Run INSOMNIQ as a service. Submit problems via API, get autonomous AI agent results asynchronously. Deploy on your own infrastructure or use the hosted version.
AGPL-3.0 · Python 3.11+ · Claude · Gemini · Ollama · uv sync and go