Recursive Sub-Agents: When Agents Spawn Their Own Children (2026)

TL;DR — Sub-agents are agents spawned by other agents. Four patterns dominate in 2026: inline tool, spawn-wait, parallel, and quality-gated. RecursiveMAS reports 8.3% accuracy gains and 35–75% token reduction when applied carefully.

The pattern

A parent agent treats a child agent as a callable. The child has its own context, prompt, and tools; the parent integrates the child's response. Recursion arises when the child can spawn its own children.

flowchart TD
  P[Parent agent] -->|spawn| C1[Child A]
  P -->|spawn| C2[Child B]
  C1 -->|spawn| GC1[Grandchild A1]
  C1 -->|spawn| GC2[Grandchild A2]
  GC1 --> C1
  GC2 --> C1
  C1 --> P
  C2 --> P
  P --> JUDGE[Quality judge]
  JUDGE -->|FAIL| P
  JUDGE -->|PASS| OUT[Result]

The four sub-patterns:

1. Inline tool — child wrapped as a tool call; parent gets a string back.
2. Spawn-wait — long-lived child, parent sends follow-ups via messages.
3. Parallel — N children spawned simultaneously (overlaps with map-reduce).
4. Quality-gated — parent specifies acceptance criteria; a separate judge evaluates child output PASS/FAIL.

When to use it

• Tasks decomposable into 2+ levels of sub-tasks.
• Codebases where some sub-tasks are themselves multi-step (e.g., "build feature X" → "design DB schema" → "write SQL").
• Coordination work too rich for flat orchestration.

CallSphere implementation

CallSphere's deepest recursion is OneRoof's listing-deep-dive flow:

• Parent — Triage Aria (supervisor).
• Child — property specialist.
• Grandchildren (spawned by property specialist) — comparable-sale finder, school-zone reader, suburb economic indicator fetcher.

The property specialist runs IMPLEMENT-THEN-VERIFY: it spawns each grandchild with explicit acceptance criteria ("return at least 3 comps with sale dates within 6 months"); a cheap judge model evaluates each grandchild's output PASS/FAIL. FAIL retries once with the critique.

Explore a live demo and compare current plans to find the right fit for your business.

Build steps with code

def run_with_judge(parent_call, criteria):
    out = parent_call()
    verdict = judge.invoke(f"Criteria: {criteria}\nOutput: {out}")
    if verdict.strip().startswith("PASS"): return out
    return parent_call(critique=verdict)  # one retry

def property_specialist(query):
    comps = run_with_judge(
        lambda critique=None: comp_finder(query, critique),
        criteria="At least 3 comps, all sale dates within 6 months"
    )
    schools = school_reader(query)
    suburbs = suburb_indicator(query)
    return synthesize([comps, schools, suburbs])

Pitfalls

• Unbounded recursion depth — cap at 3 levels. Past 3, debugging is brutal.
• Lost context across levels — children get briefs; don't pass parent's full history.
• Judge cost ignored — judges run on every child output. Use a small fine-tuned judge.
• No idempotency — retries duplicate work. Cache by (criteria, input) hash.

FAQ

Q: Inline tool vs handoff? Inline tool when the parent integrates. Handoff when the child should reply directly to the user.

Q: How deep can recursion go? Practically, 3 levels. Theoretically unbounded; in production, audit trails get unreadable past 3.

Q: Different model per level? Common — strong at the root (planning), cheaper at leaves (execution).

Q: Quality gate worth the cost? On critical paths, yes. ~30% extra cost, often 50% fewer downstream errors.

Q: How does RecursiveMAS differ? RecursiveMAS shares latent state across agents in a unified embedding, beyond just text passing. Research-grade today.

Sources

• Phil Schmid — Subagent Patterns 2026
• RecursiveMAS — arxiv
• Awesome Agentic Patterns
• MindStudio — Sub-Agents in Claude Code

Recursive Sub-Agents: When Agents Spawn Their Own Children (2026) — operator perspective

When teams move beyond recursive Sub-Agents, one question shows up first: where does the agent loop actually end? In practice, the boundary is rarely the model — it is the contract between the orchestrator and the tools it calls. The teams that ship fastest treat recursive sub-agents as an evals problem first and a modeling problem second. They write the failure cases into the regression set on day one, not after the first incident.

Why this matters for AI voice + chat agents

Agentic AI in a real call center is a different beast than a single-LLM chatbot. Instead of one model answering one prompt, you orchestrate a small team: a router that decides intent, specialists that own a vertical (booking, intake, billing, escalation), and tools that read and write to the same Postgres your CRM trusts. Hand-offs are where most production bugs hide — when Agent A passes context to Agent B, anything that isn't explicit in the message gets lost, and the user feels it as the agent "forgetting." That's why the systems that hold up under load are the ones with typed tool schemas, deterministic state stored outside the conversation, and a hard ceiling on tool calls per session. The cost story is just as important: a multi-agent loop can quietly burn 10x the tokens of a single-LLM design if you let it think out loud at every step. The fix isn't a smarter model, it's smaller agents, shorter prompts, cached system messages, and evals that fail the build when p95 latency or per-session cost regresses. CallSphere runs this pattern across 6 verticals in production, and the rule has held every time: the agent you can debug in five minutes will out-survive the agent that's "smarter" on a benchmark.

FAQs

Q: How do you scale recursive Sub-Agents without blowing up token cost?

A: Scaling comes from constraint, not capability. The deployments that hold up keep each agent narrow, cap tool calls per turn, cache the system prompt, and pin a smaller model for routing while reserving the larger model for synthesis. CallSphere's stack — 37 agents · 90+ tools · 115+ DB tables · 6 verticals live — is sized that way on purpose.

Q: What stops recursive Sub-Agents from looping forever on edge cases?

A: Hard ceilings beat heuristics. A maximum step count, an idempotency key on every tool call, and a fallback to a deterministic script when confidence drops below a threshold are what keep the loop bounded. Evals that simulate noisy inputs catch the rest before they reach a real caller.

Q: Where does CallSphere use recursive Sub-Agents in production today?

A: It's already in production. Today CallSphere runs this pattern in Healthcare and Salon, alongside the other live verticals (Healthcare, Real Estate, Salon, Sales, After-Hours Escalation, IT Helpdesk). The same orchestrator code path serves voice and chat — the difference is the tool set the router exposes.

See it live

Want to see sales agents handle real traffic? Spin up a walkthrough at https://callsphere.ai/demo or grab 30 minutes on the calendar: https://callsphere.ai/book.