Agent Pipeline

The agent pipeline is the core execution engine. When you send a message, here's what happens step by step.

1. Routing

The Router agent classifies your message into one of three categories:

Category	Route	When
Simple	Straight to executing	Quick tasks, follow-ups, complaints
Complex	Discovery, then executing	Tasks needing research, planning, external APIs
Image Generation	Straight to executing	Image creation requests

The router also detects skill matches and determines whether a skill's full instructions or just its context should be used.

Smart Classifications

• Complaints about output ("the result is wrong") route to executing, not conversational

• Alternative approach suggestions ("try using X instead") always route to executing

• Tasks producing substantial code (websites, dashboards) route through discovery

• Follow-up questions about a skill get the skill's context without its full instructions

2. Discovery (Complex Tasks)

The Planner agent researches the task before execution:

• Investigates available APIs, data sources, and approaches

• Considers constraints (free vs paid, auth requirements)

• Outlines a step-by-step plan

When Interactive Planning is enabled (Settings > Agents > Planner), the plan is proposed to you for approval before proceeding.

3. Execution

The Executor is the workhorse. It:

• Reads the user's message and any skill instructions

• Uses tools to accomplish the task (web search, browser, file operations, Zapier, etc.)

• Calls plugins when custom code is needed

• Triggers the code pipeline for substantial code deliverables

Executor Rules

1. Read the user's latest message first. User intent takes priority over skill instructions.

2. Do the task directly using tools. Only create skills/plugins when explicitly asked.

3. Use create-plugin for any substantial code output (triggers the quality pipeline).

Guardrails

The executor has several smart guardrails:

• Consecutive failure detection: Warns at 3 failures, stops at 5

• Loop detection: Breaks if the same tool is called with the same arguments 3 times

• Claim verification: Rejects responses that claim actions not backed by tool calls

• Empty-work gate: Rejects completion when no substantive work was done

• Grounding gate: Catches hallucinated URLs not found in tool results

• Search-pivot nudge: After 3+ thin search results, suggests trying a different approach

• Response-count guardrail: Forces conclusion after excessive text regeneration

4. Review

The Reviewer agent inspects the output, particularly for visual content:

• Describes what it sees (images, dashboards, HTML pages)

• Reports gaps in a structured format (REQUIREMENT | OBSERVED | GAP)

• Never suggests code fixes (that's the developer's job)

Review happens automatically for code pipeline outputs that produce visual results.

Output Routing

All user-facing output goes through the OutputRouter, which:

• Streams interim text (first iteration only, for acknowledge-then-work)

• Deduplicates repeated content

• Suppresses internal agent text (planner output is never shown to the user)

• Ensures the longest, most substantive response wins