GAINs and HCINs give you two shapes for a multi-agent workflow: the flat network and the tiered one. This is a practical guide to designing with them, updated against the contemporary, verified evidence — including the most important rule, which is knowing when not to build a network at all.

Step 0 — Should this be a network?

Start here, because the seductive failure is reaching for agents because it sounds powerful. A network multiplies token cost by roughly an order of magnitude and adds a failure surface at every seam; the largest study of multi-agent failures (MAST) found high failure rates driven by design, not model quality. Build a network only when the task decomposes into independent, parallelizable subtasks where breadth matters. For tightly-coupled, sequential work, a single strong agent with good context wins.

Step 1 — Flat (GAIN) or tiered (HCIN)?

Use a flat GAIN — one coordinator over a pool of specialists — by default. Move to a tiered HCIN only when one coordinator can't hold the task: the agent pool is large, the problem is deep, or information-handling rules require asymmetric access. Tiers buy scale at the price of more seams and more latency; don't pay it without the load.

Step 2 — Design the topology (don't freeze it)

An agent network is a graph, and its structure is a design variable. GPTSwarm shows topology choice measurably changes performance and that adaptive structures beat fixed ones; MaAS fits a topology per query at 6–45% of the cost of fixed baselines. Let the coordinator decide how many specialists, how they connect, and whether a subtask needs the network at all — rather than pouring every task through one fixed decompose-spawn-validate loop.

Step 3 — Govern the information flow

Give each agent a scoped, isolated context — only what it needs. This is HCIN's privileged/asymmetric flow, now standard practice and formally a least-privilege policy (ALARA; AgentNet's enforced asymmetric flow). Back it with a blackboard: a structured central shared memory agents post to and read from within their scope. Central memory plus per-agent read scope reduces attention dilution, error propagation, and leakage at once.

Step 4 — Verify at the seams

The failure to fear is error cascade: a wrong intermediate result gets built upon until the network reaches a confident false consensus. Place validation between stages (and between tiers in an HCIN), with diverse critics that don't share the specialists' assumptions and a requirement that every claim trace to a source. A single end-of-line gate is not enough.

Step 5 — Recruit dynamically, tear down when done

Keep the ephemerality: spin agents up for the task and dissolve them when finished. Recruitment itself can be automated — AutoAgents generates the specialists a task needs, and evolving-orchestration work selects them per round based on what the task surfaces.

The shape of a good multi-agent workflow

Decompose only when the task is parallel; pick flat over tiered unless the load demands tiers; fit the topology to the task instead of freezing it; scope every agent's context as least-privilege over a shared blackboard; verify at every seam with diverse critics; recruit dynamically and tear down. That is GAINs and HCINs as the field now understands them — the same frameworks, engineered with what the evidence confirmed.