Skip to content

Agent Orchestration

Varun S V edited this page Oct 16, 2026 · 1 revision

Agent Orchestration: T1 → T2 → T3

Cascade organizes work using a hierarchy rather than invoking a single agent for every task.

Tier Responsibility
T1 Administrator Analyze complex requests, plan sections, allocate managers and compile results
T2 Manager Decompose a section into subtasks, manage workers, verify and aggregate
T3 Worker Execute tools, solve subtasks, check work and escalate where necessary

Task-size adaptation

A simple request may be answered directly or by a single T3 worker; moderate work involves T2 and its T3 workers; complex work recruits T1 to coordinate several T2 sections. The objective is to avoid paying for elaborate orchestration when it is unnecessary.

flowchart TD
  A[User request] --> C{Complexity?}
  C -->|Simple| W[T3 worker / direct response]
  C -->|Moderate| M[T2 manager]
  C -->|Complex| P[T1 administrator]
  P --> M1[T2 section A]
  P --> M2[T2 section B]
  M --> X[T3 workers]
  M1 --> X1[T3 workers]
  M2 --> X2[T3 workers]
Loading

Dependencies, permissions and recovery

  • T1 sections and T2 subtasks use a typed task graph and DependencyScheduler. A hard failure can block dependent tasks without wasting further tokens on impossible prerequisites.
  • Workers may request bounded additional help via their manager. Plan review can pause complex runs before the work begins.
  • Verifiable criteria such as file existence or exact content can be checked mechanically before resorting to model judgment.
  • Dangerous tools still follow permission escalation T3 → T2 → T1 → user. A planned action is not automatic permission to execute it.
  • Durable checkpoints can preserve finished work across cancellation, crashes and budget exhaustion; /continue resumes where possible.

Observe and steer an execution

REPL command Purpose
/tree Live hierarchy
/comms Agent-to-agent messages
/why Complexity classification, model selection and failovers
/cost Cost, tokens and estimated delegation savings
/plan Preview a plan without running it
/replan Request another planning pass
/steer Guide currently active workers
/continue Resume a resumable run

Sources: src/core/tiers/, src/core/orchestration/.

Clone this wiki locally