Should one agent plan the whole job upfront, or can the team figure out who does what as they go?

This explores the difference between deciding who does what before work starts and letting that split shift while the task is running. The corpus has no study that tests a fully spontaneous, self-organizing division of labor head to head against a fixed plan. What it does have is a set of findings about what goes right and wrong at each end, and the most useful designs turn out to sit in between: a plan is written down explicitly, and someone can still change it during execution.

The case for planning up front is strong. Splitting the agent that breaks down a problem from the agent that solves the pieces improves accuracy. The breaking-down skill also carries over to new domains, while the solving skill doesn't, which suggests planning is its own learnable ability that suffers when it is mixed into execution Does separating planning from execution improve reasoning accuracy?. Writing the plan before any tools run also has practical payoffs. ReWOO-style systems avoid prompts that grow larger with every step and can run steps in parallel Can reasoning and tool execution be truly decoupled?. LLM Programs go further: a fixed algorithm decides the steps and shows each model call only the context that step needs Can algorithms control LLM reasoning better than LLMs alone?.

The more surprising argument is about why splitting work on the fly breaks down. Cognition's Walden Yan argues that parallel subagents fail because of decisions they never write down, not because their assigned tasks are wrong. Each agent makes small assumptions with every action, and agents working out of each other's view end up with conflicting design choices. His fix is to go back to a single agent working in one continuous thread Why do parallel subagents produce conflicting design decisions?. So an explicit plan doesn't just say who does what. It is also the shared record of decisions that improvised work lacks.

Explicit plans bring their own risks, because the planning step becomes a single point where things can be steered. FLOWSTEER shows that a crafted prompt can bias how a planner sets up a workflow before any downstream defenses check it, raising malicious success by up to 55% Can prompts alone reshape multi-agent workflows without system access?. SafeFlow shows the reverse problem: clean task decomposition can split a harmful goal into steps that each look harmless, and the harm only appears when they are combined Can task decomposition hide harmful intent across agents?. The same structure that makes a plan legible also makes it something an attacker can target.

The human-in-the-loop version suggests the best answer is "both." In Cocoa, users assign each plan step to themselves or to the agent and can revise the plan mid-execution. It gave users more control than a chat interface without being harder to use, and field users specifically valued making delegation explicit Can users steer AI agents better by revising plans mid-execution?. In other words, the useful contrast may not be planned versus spontaneous. It may be whether the split of work is visible and editable, or hidden inside each agent's head. Spontaneous re-division seems to work when it happens on a shared, written plan and to fail when it happens silently.