Agent Workflows

An agent system extends a model with tools, memory, and multi-step planning. Each step produces an output that becomes the input for the next. Failures cascade — a single bad tool call can consume all remaining context, drive up latency, and never complete the user's request.

System architecture

• Planner — decomposes task into steps
• Executor — calls tools and processes results
• Tools — APIs, code execution, search, databases
• Memory — short-term context window, optional long-term store
• Termination condition — how the agent decides it is done

What can go wrong

• Infinite loops (planner re-calls the same tool repeatedly)
• Step explosion (task decomposed into too many steps)
• Tool failure propagation (bad tool output poisons downstream steps)
• Context overflow (execution history fills context window)
• Termination failure (agent never decides task is complete)

Detect

Reliai identifies:

• execution depth increase (steps per request rising)
• latency spike (agent running longer without completing)
• tool call repetition patterns (same call appearing 3+ times)
• trace length explosion (context window consumption)

When sampling is active during an agent loop incident, evidence may be partial. Root cause will note this.

Understand

Incident example

Production research agent enters a loop, repeatedly calling a web search tool with the same query. Each iteration consumes 2,000–4,000 tokens. Requests never complete.

• Latency: 4s → 38s average
• Cost: 3x increase in token spend
• Completion rate: 91% → 58%
• Trigger: updated tool response format for search API v3

Root cause

Search API v3 changed its response schema. The planner was configured to retry if results did not contain a specific field that no longer existed in the new schema. It retried indefinitely.

Reliai identified via:

• tool call sequence comparison (v2 traces vs v3 traces)
• detection of repeated tool call signatures
• prompt diff showing unchanged termination condition

Fix

• Update planner termination condition to handle missing field gracefully
• Add max step limit (hard cap at 12 steps)
• Update tool output parser for API v3 schema

Prove

Key takeaway

Agent failures are systemic, not single-response errors.

A single bad tool output can cascade across every step of every request in flight. The only way to detect loop patterns is trace-level analysis — log lines and error rates alone won't show you the execution structure.