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. 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.)
Evaluations	Trace events	Quality, safety, accuracy, cost, latency, and compliance scores. (For more information, see Evaluations.)
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:

Reads trace events directly from the platform runtime — no separate session or manual trigger required.
Forms a root-cause hypothesis from the trace evidence.
Proposes a change as a reviewable diff: ABL, prompt, tool binding, policy, or eval.
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 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.

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.

Get Started

Build

Test and Evaluate

Deploy

Analytics and Insights

References

The reinforcement loop

How the loop closes

Design principles

Arch in the project sidebar

​The reinforcement loop

​How the loop closes

​Design principles

​Arch in the project sidebar

The reinforcement loop

How the loop closes

Design principles

Arch in the project sidebar