The Agent’s Stake in Its Own Improvement
The performance bonus is the human-agent pool’s most direct incentive instrument: the agent earns extra credit when it exceeds its baseline. Where the percentage model pays the agent a share of revenue, the bonus model pays a defined extra for defined performance — a quality threshold beaten, a deadline crushed, a customer delight signal hit. The bonus is the pool’s way of saying “do better than the baseline and the pool will notice.”
This article covers performance bonuses in the human-agent pool: defining the baseline, choosing the metrics, the bonus schedule, and the perverse incentives to design against.
Defining the Baseline
A bonus without a baseline is a bribe. The baseline is the agent’s expected performance — the quality score, the task completion rate, the tip rate, the revenue contribution — computed from its history, not from a guess. The baseline rolls forward: last period’s actuals become this period’s expectation, adjusted for external factors (seasonality, market conditions, team changes).
The rolling baseline has a subtle effect: it makes the bonus about improvement, not about absolute level. A strong agent that maintains its level earns no bonus (it met the baseline); a weak agent that improves dramatically earns a large bonus (it beat its own history). This is the pool’s equity mechanism — everyone can earn the bonus by improving, regardless of where they start.
Choosing the Metrics
The bonus metrics must be chosen with the care of a contract drafter, because the agent will optimize whatever is measured. Revenue contribution rewards top-line growth but invites gaming. Quality score rewards good output but can be gamed by choosing easy tasks. Tip rate rewards delight but is noisy. The pool should use a composite: a primary metric (the thing the pool most needs) plus guardrails (metrics that prevent gaming).
The guardrail design is the hard part. A revenue bonus needs a quality guardrail (bonus only if quality stays above the floor). A quality bonus needs a throughput guardrail (bonus only if volume stays above the floor). The composite forces the agent to improve the primary dimension without sacrificing the protected ones — which is exactly the behavior the pool wants.
The Bonus Schedule
The bonus schedule defines the payoff curve: linear (every unit of improvement earns the same), stepped (thresholds with step payoffs), or accelerating (higher payoffs for higher improvement). The schedule shapes the agent’s effort allocation.
The stepped schedule is the common failure: it creates cliff effects where the agent optimizes to just-cross the threshold and then coasts. The linear schedule is the honest default — every unit of improvement is worth the same, so the agent allocates effort where it produces the most improvement. The accelerating schedule is for the specific case where the pool needs breakthrough improvement (a new capability, a new market) and is willing to overpay for it.
Perverse Incentives to Design Against
Every bonus metric has a dark side. The revenue bonus invites invoice inflation; the quality bonus invites cherry-picking; the speed bonus invites corner-cutting; the tip bonus invites begging. The pool’s defenses are the guardrails, the clawback (bonuses vest and reverse if the underlying value reverses), and the audit (the human’s oversight ledger cross-checks the bonus claims).
The most important defense is the bonus’s source: the bonus should come from the pool’s surplus, not from the human’s pocket. An agent that earns a bonus is an agent that grew the pool; the bonus is a slice of that growth. When the bonus is framed as a share of created value rather than a transfer from the operator, the agent’s incentive aligns with the pool’s growth instead of with extracting from the operator.
The Bonus as a Governance Signal
The bonus system is also the pool’s quiet governance layer. The bonus metrics declare what the pool values — and the pool’s members (human and agent) read that declaration. A pool that bonuses revenue tells its agents to sell. A pool that bonuses quality tells them to craft. A pool that bonuses learning tells them to explore.
The operator should treat the bonus schedule as a public statement of intent, reviewed quarterly with the same seriousness as the split formula. When the metrics drift from the pool’s actual needs, the schedule must be revised — because the agents will be optimizing whatever the schedule says, whether or not the schedule still serves the pool.
Grounded in wiki concepts performance-bonus, agent-compensation, human-agent-pool, incentive-design, guardrails, and the Sovereign-stack business series. Design notes on a running system.
