Product Thesis & Experience
Relevant source files
The following files were used as context for generating this wiki page:
The Noesis product thesis centers on the concept of autonomous compounding, where an agentic system becomes demonstrably better through normal use without requiring the user to manually operate a learning pipeline. The core claim is that later work is better because earlier work happened, driven by an evidence-governed loop that transforms ordinary interactions into capability revisions.
The user experience is designed to be "magic on the surface; evidence underneath." Noesis operates with a conversational-first interface, providing immediate value through a "foreground turn" while performing ambient reflection, evaluation, and capability authorship in the background. Users maintain sovereignty through high-leverage controls—such as pinning, vetoing, and reverting adaptations—ensuring that while the system evolves autonomously, it remains inspectable and corrigible.
Sources: noesis-autonomous-compounding-implementation-plan.html:158-164, compounding-partnership-product-loop.html:200-213, AGENTS.md:5-13
Core Product Thesis
The thesis is built on three pillars that balance LLM-driven semantic work with strict, code-enforced rails.
Compounding Advantage
The system must produce a measurable advantage where an activated revision changes a later relevant turn. This compounding effect is validated through Effect-free paired shadow replay, which compares a served revision against its baseline to ensure the adaptation provides genuine utility rather than just "inject-everything" accumulation.
Sources: noesis-high-leverage-correction-plan.html:77-105, AGENTS.md:6-7
Evidence-Governed Learning
Every durable learning must earn its place. The system begins at the narrowest plausible scope and requires recurring evidence before broadening. This prevents the system from generating global rules based on isolated repository-specific corrections.
Sources: compounding-partnership-product-loop.html:439-447, AGENTS.md:8-9
Human Sovereignty
The architecture enforces a clear boundary between generated content and protected authority. While an LLM can author prompt revisions and tool code, it cannot self-modify permission manifests, activation pointers, or evaluation results. The user remains the final authority, able to contest or undo any adaptation.
Sources: noesis-autonomous-compounding-implementation-plan.html:237-243, AGENTS.md:12-13
The Compounding Loop
The product experience follows a strict lifecycle that moves from signal detection to served behavior. This loop ensures that all adaptations are grounded in real-world evidence.
Learning Lifecycle Flow
The following diagram illustrates the flow from a user's initial interaction to the deployment of a new system capability.
The Noesis autonomous loop ensures that useful work is completed in the foreground before ambient reflection begins the background learning process.
Sources: noesis-autonomous-compounding-implementation-plan.html:171-180, compounding-partnership-product-loop.html:362-373
Reflection and Capability Revisions
Reflection is triggered by specific events: explicit corrections, recurring failures, surprising successes, or direct user requests. If reflection yields a hypothesis, Noesis authors a CapabilityRevision, which couples prompts, skills, and tools into a single unit.
Sources: noesis-autonomous-compounding-implementation-plan.html:366-380, AGENTS.md:32-34
Collaboration Postures
Noesis avoids manual "modes" by inferring collaboration posture dynamically based on the user's request. This inference is conservative and corrigible.
| Posture | Intent | Default Heuristic |
|---|---|---|
do_for_me | Explicit execution | Clear, bounded execution or patch requests. |
work_with_me | Intellectual partnership | Ambiguous, creative, or research-oriented requests. |
Conversational Override: Explicit user instructions (e.g., "Just do it" or "Teach me instead") immediately override the system's inferred posture and are cited in the turn's metadata.
Sources: compounding-partnership-product-loop.html:400-420, AGENTS.md:7-8
User Experience and Controls
The interface prioritizes low-noise ambient activity, keeping the chat uncluttered while providing deep transparency through expert commands.
Ambient Status
Learning activity appears as compact, non-blocking status lines below the chat input:
learning · reflecting on 2 correctionslearned · improved source-grounded research
Sources: noesis-autonomous-compounding-implementation-plan.html:619-625
Expert Command Set
Users can use expert shortcuts to audit and control the system's learning state.
| Command | Description |
|---|---|
/why | Explains the active turn, active revision IDs, and citations. |
/changes | Lists recent experiments, revisions, and activations. |
/criteria | Lists, adds, or pins user-defined evaluation criteria. |
/revert | Restores the prior active capability revision immediately. |
/approve | Manually approves a candidate that requires expanded authority. |
| Sources: noesis-autonomous-compounding-implementation-plan.html:636-655, compounding-partnership-product-loop.html:565-575 |
Causal Metrics
To ensure the thesis of compounding advantage is falsifiable, Noesis tracks four central metrics through shadow replay.
Metrics are derived from paired shadow replays, comparing the system's performance with and without a specific adaptation.
Sources: noesis-high-leverage-correction-plan.html:108-144, noesis-dogfood-protocol.md:38-47
Metric Definitions
- Served-revision win rate: Paired wins of the served revision over its baseline. If this is at chance, the adaptation is mere "decoration."
- Scope leakage rate: The share of turns where an adaptation was served but was unrelated to the actual task.
- Context tax: Separates marginal served tokens from baseline tokens to ensure advantage isn't simply bought with bloated prompts.
- Correction recurrence: Measures if a specific user dissatisfaction signature falls after an adaptation is activated.
Sources: noesis-high-leverage-correction-plan.html:125-144
Conclusion
The "Product Thesis & Experience" of Noesis establishes a framework for a self-improving agent that respects human agency. By grounding every change in evidence-linked hypotheses and measuring success through causal metrics, Noesis aims to prove that LLM-driven systems can compound their utility over time without sacrificing reliability or user control. The experience remains conversational and magic for the end-user, while maintaining a strict, auditable, and reversible architecture for the developer.
Sources: noesis-autonomous-compounding-implementation-plan.html:815-820, compounding-partnership-product-loop.html:639-650