When you spawn a child, it either inherits the parent's whole mind or starts from a clean brief. The choice trades cache savings and nuance against bias and an injection surface.
Why this, for you: this is a per-spawn decision you make constantly, and the wrong default is
expensive in a way that's invisible until it bites — a forked code reviewer rubber-stamps the bug it should
catch, because it remembers why every line was written. Knowing which to pick is the difference between a real
second opinion and a confirmation echo.
A forked subagent inherits the parent's entire system prompt, tools, and message
history. A fresh subagent starts with only the brief the orchestrator constructs. Same axis exists in
every harness that spawns children — and it's chosen per task, not globally.
1 The decision table
Spawn fresh when…
Spawn a fork when…
The child must question a parent decision (review, audit, adversarial test)
The child must extend a parent decision (design variation, continuation)
The parent touched untrusted content and the child has egress
The parent's reasoning is load-bearing and a brief would be lossy
The task is one-off — no siblings to amortize cache warmup
Two or more siblings will run from the same starting point
The child must reset bias to disagree
The child must preserve nuance to agree intelligently
2 Why forks are cheap — and why that cuts both ways
A fork's first request shares the parent's prefix exactly — same system prompt, tools, and history. The
prompt cache matches on exact prefix, so the fork reads from cached tokens and bills only the appended fork directive.
Cache reads run at roughly 10% of the standard input rate. A fresh named subagent has a different
prompt and tool set, so its prefix doesn't match — no cache hits, and it warms its own cache from scratch.
The mechanism cuts both ways. Forks are cheap precisely because they carry the parent's
entire input distribution — which is also why they inherit its biases, blind spots, and accumulated tool
results. Fresh subagents reset that distribution; that reset is what makes them useful for adversarial work.
3 Where forking backfires
Forking a review or audit. The fork remembers why every decision was made, and confirmation bias rubber-stamps it. In a direct test, a forked reviewer returned cosmetic notes on auth code; a fresh subagent on the same code flagged a missing constant-time token comparison — a real security bug.
Forking a single small task. A one-off fork on a heavy (e.g. 180k-token) parent pays the cache-write tax with no siblings to amortize against. Forks earn their keep when batched.
Forking a trifecta-sensitive child. A fork pulls in every accumulated tool result — web fetches, MCP output — and drops the input isolation subagents otherwise provide. If the parent has any lethal-trifecta exposure, the fork inherits the injection surface.
Forking past the context cliff. Forks copy the entire parent window, so a degraded parent hands the fork a degraded baseline — propagating bloat rather than solving it.
Enabling fork mode changes two things at once
In Claude Code, CLAUDE_CODE_FORK_SUBAGENT=1 turns general-purpose spawns into forks and runs
every spawn in the background. Named subagents still spawn fresh regardless. The manual /fork command
works either way — the variable governs automatic forking. A fork cannot spawn further forks.
↪ Your win: fresh to disagree, fork to extend
Default fresh for reviews, audits, and any child that must disagree — a fork anchors on the parent's bias.
Fork when the parent's reasoning is load-bearing and 2+ siblings run from the same start.
Forks share the parent's cache — first request bills at ~10% read rates because the prefix matches.
Never fork a child with egress when the parent touched untrusted content — the fork drops input isolation.
It's per-task, not global — even with fork mode on, named subagents still spawn fresh.
Retrieval practice — recall, don't peek
Question 1For a security audit of code the parent just wrote, you should spawn…
Question 2A fork's first request is cheap because its prefix…
Question 3A fork is dangerous for a child with egress because it…
Question 4The fork-vs-fresh choice is made…
Question 5 · spaced recall from Lesson 3Fan-out synthesis collapses when the N agents…
Ask me anything. Want a rule of thumb for when a brief is "too lossy" to skip forking, or how the
cache math plays out on a heavy parent with several siblings? Next in Part 2: Handoffs and Coordination
Contracts — passing work between agents without losing it.