“Fork and spawn both create separate working contexts, but they behave differently. The test is: would the child do better seeing what the parent decided? If yes, fork. If no, spawn.” — Lou
Session context: 2026-06-11_Mastermind — Lou’s 25-minute “ambient intelligence” walkthrough crystallized a distinction he’d been circling for weeks: there are two ways to hand work to a sub-process, and choosing wrong either contaminates the result or starves it.
Core Idea
Both a fork and a spawn run work in an isolated context and return only a result — neither pollutes your main conversation with intermediate drafts, tool calls, or reasoning. The difference is what the child starts with.
A fork inherits the parent’s context. It’s a virtual copy of where you already are — the child knows what the parent already decided, then does its work and returns only its high-signal output. Use it when inherited context helps: the parent has already chosen the angle for an article, and the draft-writer should build on that choice rather than re-litigate it.
A spawn starts cold. It gets only the parameters you hand it and nothing else. Use it when inherited context would contaminate the work. The canonical case is adversarial review: you don’t want the reviewer to inherit the parent’s confidence in the current direction. You want a colder read — someone who hasn’t already been persuaded.
So the choice isn’t about speed or cost; both isolate context equally. It’s about contamination versus continuity. One question settles it: would this child do better seeing what the parent already decided, or worse? Continuity → fork. Independence → spawn. This is the design decision that sits underneath every multi-stage AI workflow, and most people never name it — they just dump everything into one conversation and pay for all of it on every subsequent turn.
The payoff compounds. In a forked pipeline, a writing job that burned 350K tokens across five drafts returns only the final artifact to your context. You can re-run your last query — “didn’t like it, try again” — and the whole pipeline re-executes without you scrolling back through stale intermediate work to reconstruct state.
Practical Application
Before delegating any step to a sub-agent or skill, ask the one diagnostic question out loud: “Would this step do better knowing what I’ve already decided, or would that bias it?”
- Fork drafting, revision, formatting, and any step that should honor prior decisions.
- Spawn adversarial review, red-teaming, independent estimates, and any step where you want a fresh judgment uncontaminated by the parent’s momentum — the same logic behind running isolated agents instead of a debate when the goal is discovery.
Then require the child to return only its conclusion (“don’t send back your thinking, pros, or cons — just the final answer or the artifact path”). That single instruction converts an isolated process into a high signal-to-noise result.
Related Insights
- Insight - Forked Skills as Context Isolation — Run Sub-Agents Without Polluting Your Conversation — established forking as context isolation; this adds the spawn counterpart and the decision rule between them.
- Insight - Skill Orchestration — Forking, Quality Gates, and Pipeline Mechanics Make Chains Actually Run — fork/spawn is the execution primitive those pipelines are built from.
- Insight - Isolation Outperforms Debate When the Goal Is Discovery, Not Refinement — the “spawn cold for independent judgment” principle, applied to multi-agent exploration.
- Insight - Model Altitude — Route Model and Effort by Workflow Step, Not by Whole Artifact — once you’ve chosen fork vs spawn, you still choose which model and effort the child runs at.
- Insight - Capability-Architect — Compile a Workflow Into an Inheritable Skill Bundle — the architect decides fork-vs-spawn for each step as part of compiling a capability.
Evolution Across Sessions
Builds on Insight - Forked Skills as Context Isolation — Run Sub-Agents Without Polluting Your Conversation (2026-01-15), which established that you can fork a skill to keep its work out of your conversation. The new development: spawn is the other half of the pair, and the choice between them is a deliberate design decision governed by a single question about context inheritance. This sharpens Insight - Isolation Outperforms Debate When the Goal Is Discovery, Not Refinement (2026-04-23) into a per-step rule rather than a whole-workflow stance.