> ## Documentation Index
> Fetch the complete documentation index at: https://koreai-content-gov.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Optimize with Arch AI

After your agents go live, Arch reads production traces, runs evaluations, and proposes improvements as reviewable patches — continuously, without manual triage. Each iteration raises the baseline quality across the entire project.

## The Reinforcement Loop

Deployment is not the end of the pipeline — it is the start of a continuous cycle. Production evidence feeds evaluations, Arch analysis, reviewable patches, and promotion gates. The platform harness stays constant while models, tools, and workflows improve underneath.

```mermaid actions={false} theme={null}
flowchart LR
    A[Production Traces] --> B[Evaluations]
    B --> C[Arch Analysis]
    C --> D[Reviewable Patch]
    D --> E[Promote]
    E --> A
```

Each stage produces a defined output that feeds the next:

| Stage                 | What Flows In       | What Flows Out                                                                                                                        |
| --------------------- | ------------------- | ------------------------------------------------------------------------------------------------------------------------------------- |
| **Production Traces** | Live agent sessions | Conversation, handoff, workflow, tool, guard, and retrieval events. (For more information, see [Sessions](/agent-platform/sessions).) |
| **Evaluations**       | Trace events        | Quality, safety, accuracy, cost, latency, and compliance scores. (For more information, see [Evaluations](/agent-platform/evals).)    |
| **Arch Analysis**     | Eval results        | Root-cause hypothesis and proposed improvement lever.                                                                                 |
| **Reviewable Patch**  | Arch analysis       | ABL, workflow, tool, prompt, policy, or eval diff.                                                                                    |
| **Promote**           | Approved patch      | Deployment behind gates, canary, or rollback.                                                                                         |

Each approved patch is deployed to production, generating new traces that feed the next iteration.

### How the Loop Closes

When Arch detects a degradation, a coverage gap, or a routing inefficiency, it:

1. Reads trace events directly from the platform runtime — no separate session or manual trigger required.
2. Forms a root-cause hypothesis from the trace evidence.
3. Proposes a change as a reviewable diff: ABL, prompt, tool binding, policy, or eval.
4. Submits the patch for compiler validation before a developer reviews the diff.

Developers must approve every change. The compiler validates the structure of each proposal before it reaches review, so broken references and incomplete policies are caught before promotion.

### Design Principles

The reinforcement loop is designed around three characteristics that make continuous improvement sustainable at scale.

| Loop Characteristic     | Description                                                                                                                                    |
| ----------------------- | ---------------------------------------------------------------------------------------------------------------------------------------------- |
| **Autonomous analysis** | Arch turns signal into concrete improvement proposals. Humans set the bar and retain approval — the analysis loop runs without manual triage.  |
| **Customer-gated**      | Every evaluation and promotion gate is domain-specific. You define what "good" means; the platform enforces the same process for every change. |
| **Compounding**         | Each run enriches the evaluation suite. Each evaluation sharpens the next optimization. Each patch strengthens the reusable artifact.          |

## Arch in the Project Sidebar

Arch lives in the project sidebar — available throughout the agent lifecycle, not just during authoring. From the sidebar, you can monitor project health, act on recommendations, and trigger workflows without context-switching.

<img src="https://mintcdn.com/koreai-content-gov/u2TE6zU9j_201tS5/agent-platform/images/arch-ai/arch-ai-continuous-improvement.png?fit=max&auto=format&n=u2TE6zU9j_201tS5&q=85&s=423fc3b411dc4cb13298ab4d5ffc0c55" alt="Project health monitoring and continuous improvement by Arch" width="1948" height="1206" data-path="agent-platform/images/arch-ai/arch-ai-continuous-improvement.png" />

| Feature                  | What It Does                                                                                                        |
| ------------------------ | ------------------------------------------------------------------------------------------------------------------- |
| **Project Health**       | Runs continuous status checks and flags specific warnings when anomalies are detected.                              |
| **Suggested Workflows**  | Surfaces contextual next steps — Add Agent, Debug Issue, Improve, Review Topology — based on current project state. |
| **Quick Actions**        | One-click access to Explore project, Health check, and Review integrations.                                         |
| **Auto Recommendations** | Proposes concrete fixes grounded in your actual project, such as missing session memory on a stateful agent.        |
| **Contextual Help**      | Grounds every suggestion in your project — not generic guidance.                                                    |