What makes delegation work beyond just splitting tasks?

Current multi-agent frameworks treat delegation as task decomposition: split a goal into subtasks, assign them to agents, collect results. This is necessary but insufficient. Delegation in the full sense involves the transfer of authority, assignment of responsibility, implication of accountability for outcomes, risk assessment moderated by trust, capability matching, continuous performance monitoring, dynamic adjustments based on feedback, and ensuring completion under specified constraints.

The framework identifies 11 task characteristic axes that determine how delegation should be designed:

1. Complexity — difficulty level, correlated with sub-steps and reasoning sophistication
2. Criticality — importance and severity of failure consequences
3. Uncertainty — ambiguity regarding environment, inputs, or success probability
4. Duration — time-frame from instantaneous sub-routines to weeks-long processes
5. Cost — computational expense including token usage, API fees, energy
6. Resource requirements — specific tools, data access, human capabilities needed
7. Constraints — operational, ethical, or legal boundaries limiting the solution space
8. Verifiability — difficulty and cost of validating outcomes; high verifiability (formal proofs, code verification) enables "trustless" delegation; low verifiability (open-ended research) requires high-trust delegatees or expensive oversight
9. Reversibility — whether effects can be undone; irreversible tasks (financial trades, database deletion) require stricter liability firebreaks than reversible tasks (drafting an email)
10. Contextuality — volume and sensitivity of required external state; high-context tasks introduce larger privacy surfaces
11. Subjectivity — whether success criteria are preference-based or objective; highly subjective tasks require human value specification and iterative feedback loops

Four design variables complete the framework: granularity (fine-grained vs coarse-grained objectives, where coarse requires further decomposition by delegatee), autonomy (full autonomy vs prescriptive specification), monitoring (continuous, periodic, or event-triggered), and reciprocity (whether delegation is one-way or mutual in collaborative agent networks). Reciprocity matters because in multi-agent systems, agents may delegate to each other — the delegation relation is not always hierarchical.

Verifiability determines whether evaluation is possible at all. The verifiability axis is not only a delegation criterion — it is a precondition for evaluation itself. Subjective, non-verifiable work (aesthetic judgment, strategic framing, interpretive analysis) leaves AI performance assessable only through proxies: style, fluency, apparent comprehensiveness, polish. Since Does polished AI output trick audiences into trusting it?, these proxies are exactly the dimensions AI is optimized to satisfy. The result is structural: as AI handles more subjective work, evaluation becomes harder rather than easier, because the only signals available to the evaluator are signals the system was trained to maximize. Delegation on low-verifiability axes is therefore not just a trust problem — it is a measurement problem where the instruments available all read positive.

The most design-relevant axes are verifiability, reversibility, and subjectivity because they determine the delegation contract — not just what to delegate, but how much trust, oversight, and rollback capacity the system needs. Since Does structured artifact sharing outperform conversational coordination?, the delegation framework provides the principled basis for what those SOPs should specify.

Source: Agents Multi Architecture