A tool called qm appeared on Hacker News this week. It is a multiplayer agent harness — software that lets multiple AI agents divide up tasks, hand off work to each other, and complete multi-step projects without a human in the loop at every stage. The repository is sparse and early. The underlying pattern it represents is neither sparse nor early.

Multi-agent coordination is the capability that closes the last gap most white-collar workers thought protected them: complexity. A single AI can draft an email. A team of AI agents can research a vendor, draft the proposal, check the contract terms, flag the compliance risk, and schedule the follow-up meeting. That sequence describes a workday for a lot of households that earn $60,000 to $120,000 a year.

What is actually changing

Until recently, AI automation threatened tasks, not jobs. The argument from labor economists was that automation disrupts specific duties — data entry, invoice coding, basic drafting — but that jobs are bundles of tasks, and humans would keep the bundle even as individual pieces got automated away.

Multi-agent systems challenge that framing. When agents can hand off between each other, the bundle itself can be automated. Recent BLS occupational data shows that administrative support, paralegal work, junior financial analysis, and entry-level software QA are the roles with the highest proportion of tasks that fit neatly into agent-executable workflows. These are not marginal jobs. They are the jobs that fund mortgages and college savings accounts in a large share of two-income households.

The shift is not instant. Enterprise procurement cycles are slow. Liability questions around autonomous AI action are unresolved. But the directional pressure is clear, and a family that waits for a layoff notice to start thinking about this is starting too late.

The honest uncertainty here: nobody has reliable data on the pace of displacement. Projections from think tanks range from "modest disruption over a decade" to "significant displacement within three years." The range itself is the signal. This is not a stable, predictable transition.

What we would actually do

Audit which parts of your job an agent could already do. Spend 30 minutes writing down your last five workdays in task-level detail. Categorize each task: requires physical presence, requires relationship trust, requires licensed judgment, or requires only information processing. The last category is the exposure. Knowing your ratio is more useful than generalized anxiety about AI.

Jobs are bundles of tasks, and the proportion of information-processing tasks in yours determines how substitutable you are. If more than half your week falls into that last bucket, that is not a reason to panic but it is a reason to deliberately shift toward the other three — client-facing roles, licensed functions, physical-presence work — before the choice is made for you.

Build one income stream that does not depend on an employer's headcount decisions. This does not mean quitting your job. It means spending three to five hours a week developing a skill, service, or small revenue source you own. Recent Federal Reserve household survey data consistently shows that households with even modest secondary income weather job disruptions significantly better than those with a single income source.

A side income of $500 a month is not retirement. It is a pressure-release valve: it buys you months of negotiating room if a layoff comes, and it builds the habit of working outside a single employer's ecosystem.

Get specific about your field's hiring pipeline, not the economy in general. Aggregate unemployment numbers will mask sector-level disruption. If you work in paralegal services, junior accounting, or content moderation, look at job-posting volume in your specific niche on a quarterly basis. A 20% drop in posted roles in your field over six months is a more actionable signal than any national statistic.

Prioritize credentials that require supervised hours or a physical license. Electricians, nurse practitioners, licensed appraisers, commercial drivers — these roles share a common feature: an AI cannot legally sign off on the output. That regulatory moat is not permanent, but it is durable on a five-to-ten-year horizon, which is the planning window that matters for most families with children at home.

The bigger picture

The qm repository on GitHub is not the thing to watch. It is one visible symptom of a capability curve that has been climbing steadily. The households that come through this transition intact will not be the ones who predicted the exact timing. They will be the ones who treated income like infrastructure — redundant, diversified, and regularly audited for single points of failure.

Durability is the goal. Not catastrophe avoidance, not getting rich off disruption. Just a household that can absorb a bad quarter and keep moving.