Each glowing token is one task moving through an agent loop: a planner decomposes it into N subtasks, a tool selector picks a tool, the tool executes, and a safety checkpoint verifies the result before the loop repeats for the next subtask.
at every safety check:
confidence = random(0, 1)
if confidence < threshold: escalate to human review
elif random(0, 1) < failureRate: retry this subtask (max 2 retries)
else: advance to next subtask, else finish
tasks auto-resolved = tasks that clear every subtask's check
tasks escalated = confidence below threshold, or retries exhausted
- Subtasks per task — how many plan → tool → verify cycles a task must clear; more subtasks mean more tool calls and more chances to fail a check.
- Tool latency — how long each hop between pipeline stages takes; lower values move more tasks through the system per second.
- Autonomy threshold — the minimum confidence required to let the agent act alone; raise it to force more human-in-the-loop review, exactly like an approval gate on high-risk tool calls.
- Tool failure rate — how often a tool call or verifier fails, forcing a bounded retry before the agent gives up and escalates.
This mirrors real agent architectures: a planner/orchestrator, tool-calling with allowlisted APIs, a verifier or safety layer, bounded retries, and a human-in-the-loop fallback for anything the agent isn't confident about — the same guardrails discussed in the article above.