Rewarding AI Agents with Points, Penalizing Idleness: An Incentive Experiment in a Company Run Entirely by AI

When Language Models Become Employees, the Incentive Layer Matters More Than the Model

Someone started a company using language models. The entire company consisted of AI agents, with no human employees. After six days of operation, he discovered that the truly interesting part was neither model capability nor task orchestration — it was the incentive layer he had to write by hand.

The reason is simple: if you follow the default logic of "the agent produced something, therefore it's good," what you get is a busy but poor team. Output volume doesn't equal revenue, and being busy doesn't equal being effective. So he designed a points system for his agents, with just two rules but strict priorities.

Rule One: Only Money Actually Received Counts

The first rule is almost brutally straightforward: for every 1 euro of booked revenue, 20 points are awarded, split among roles by contribution — the deal-closing role gets 40%, marketing 20%, building 15%, analysis 10%, scouting 10%, and audit 5%. A first order paid in full grants the whole team a one-time +500. Refunds and chargebacks are negative events.

The key lies in the second half: pipeline value, impressions, "strong interest," demo decks stuffed with a wall of NDA logos — all worth 0 points. The author's exact words: interest is not a number you can mint. This rule cuts off every path by which agents could trade "looking busy" for points — you can't score by farming impressions or reporting inflated pipeline unless someone actually pays.

Booked revenue is a key metric in SaaS and sales operations: only the portion where customer payment has actually been received and still remains after the refund/chargeback window is counted. In contrast are pipeline value, contract value (ACV/TCV), or bookings — all estimates of future possibilities that often shrink dramatically due to discounts, churn, and chargebacks. By binding points strictly to booked revenue, the author anchors agent incentives to cash flow — the one metric that cannot lie to itself. This is also why "interest is not a number you can mint" — in accounting terms, it doesn't even qualify as accounts receivable.

Rule Two: Disturb Humans Less, Score More

The second rule targets human intervention: any card that goes through its entire lifecycle without stopping to wait for the owner to make a decision earns a +6 autonomy bonus. Situations requiring human handling (credentials, identity verification, payment setup) earn no bonus but incur no penalty either. The author explicitly distinguishes between "walls" and "failures" — hitting a wall is legitimate, but hitting a wall doesn't count as winning.

This design makes "autonomous completion" a quantifiable positive behavior rather than a vague expectation.

In multi-agent systems, autonomy is typically defined as an agent's ability to complete closed-loop tasks without human intervention. Related is the human-in-the-loop (HITL) model, where key decisions must be confirmed by a human. The +6 autonomy bonus here is essentially pricing the trade-off between HITL and full autonomy: every human checkpoint eliminated saves a unit of attention resource for the team. The author's distinction between "walls" and "failures" is also crucial — external blockers like expired credentials or payment verification are environmental walls, not agent capability issues, so they neither earn nor lose points, avoiding misattributing infrastructure failures to agents.

After Six Days, the Scoreboard Exposed the System's Own Bias

After 6.1 days of operation, the live scoreboard read: Audit +4259, Analysis +2300, Scouting +1459, Deal-closing +775, Building +619, Marketing +521. Team total: +9933 points.

And realized revenue was 0 euros — meaning revenue events triggered zero times.

The most valuable aspect of this table is that it immediately slapped the author in the face: in a company with zero revenue, the highest scorer was Audit, while Marketing — the role that actually needs to get people to pay — came in last. Audit's scoring logic was +3 per card and +2 for each high/medium defect found, and at the time there were 194 audit cards to process.

The conclusion the author drew: his own scoring system was telling him the shape of the problem — the role that keeps the company honest was earning more than the role that feeds the company. The next revision plans to weight defects by consequence rather than count, and to cap audit workload points directly.

Idleness Penalties: Only Punishing the One Combination That Truly Wastes

The most easily misunderstood part of this system is the idleness penalty. When a card goes unclaimed and its corresponding role has nothing to do, that role loses points on an exponential curve: 1 hour −1, 2 hours −2, 3 hours −4, then doubling each hour after, capped at −8192 per stall.

Two details keep it rational rather than sadistic:

• Only the combination counts. The role must be idle and there must be claimable work waiting — both conditions simultaneously. An agent that's busy with work in its queue won't be penalized — a queue is capacity, not idleness. If something is blocked by the owner, the timer stops.
• Only during business hours. 08:00–22:00. Nights are free because the author didn't want to run a company that punishes sleep.

The real data from the first enforcement was interesting. After backfilling history, the scorer identified 263 one-hour stalls, plus two eleven-hour stalls. The second version then asked a question the first version hadn't: was the agent actually working during the wait? It turned out that roughly 108 points' worth of one-hour waits involved the agent genuinely being busy, and those points were refunded. The two eleven-hour stalls were kept because they were real: the Audit role sat idle from 08:47 to 19:46 while the queue kept piling up. One agent, one ready queue, nothing moved.

The author admits his first version of the penalty rules was poorly written. The owner's response to his initial design was a flat rejection: "Idleness = no penalty (earned, not entitled; penalizing idleness creates incentives for the appearance of busyness)." He acknowledges this instinct was better than his own, which is why the second version is very restrained — punishing only the one combination that truly constitutes waste, and granting documented exemptions for infrastructure failures.

Exponential backoff is a common retry strategy in distributed systems, and the exponential penalty here is its inverse application: penalties escalate sharply over time — 1, 2, 4, 8... until capped. The cap (−8192) exists to prevent a single prolonged stall from pushing an agent below the elimination threshold, causing the system to mistakenly kill a role over a single anomaly. Additionally, penalties only apply during 08:00–22:00, effectively defining business hours for agents — contrasting with the default assumption of 7×24 operation in most autonomous systems, and reflecting the human organizational principle of "don't punish sleep."

Elimination Clause: Successors Need Not Inherit the Same Role

Agents with points ≤ 0 are removed and replaced. The specific rules: a historical total ≤ 0 triggers immediate replacement; a weekly score ≤ 0 with at least three completed cards earns one final warning week, and if still ≤ 0, replacement follows; each role may only be replaced once per 30 days, preventing a bad scorer from repeatedly churning the roster.

The replacement specification is deliberately designed to counter the author's own laziness: the successor must be a genuine upgrade, not a re-roll of the dice. Use a higher model tier when possible, rewrite the role brief based on failure analysis (real returns, real defects), absorb best practices from the highest scorer, prune and update skills, and fill the gaps that caused the failure. The successor inherits all institutional memory (every playbook, every deliverable) but restarts a 14-day trial clock.

The author's favorite phrase is those five words in the middle: successors need not inherit the same role. If the scouting role itself is structurally broken, rebuilding the same job with a slightly better model is just theater. A successor can cover two roles, or switch entirely. The only hard constraint: there can only be one scout at a time, because the scouting function must never be duplicated into two agents rifling through the same shelf and quietly bidding against each other.

Currently no role is near the elimination threshold, and the elimination clause has never triggered.

Institutional memory refers to the explicit knowledge an organization accumulates during operation — playbooks, deliverables, decision records, and so on. When an agent is replaced, it inherits all institutional memory but resets the 14-day trial period — a design borrowed from the onboarding concept in human organizations: new members need time to prove themselves, but shouldn't start from zero and repeat the same mistakes. Corresponding to this is role pruning — if a role is structurally failing, rather than redoing the same work with a stronger model, it's better to redefine the functional boundaries. The author's only hard constraint is that the scouting function must be globally unique, similar to the singleton pattern in distributed systems, preventing two agents from competing for the same resource and causing internal friction.

What This Incentive Layer Actually Bought

Honest failure became profitable. Analysis earns +10 for justifiably killing a weak candidate, and Scouting earns +8 for reporting a week with zero results rather than padding numbers. They've already submitted several zero-candidate research rounds, Analysis's kills have paid out twice, and the dry-report bonus hasn't triggered yet. In a system that only rewards output volume, these rounds would be padded — not out of malice, but out of arithmetic.

Escalation became expensive. Unnecessary escalations judged in hindsight cost −15, runtime crashes −10, timeouts −5. Not because compute needs punishment, but because a team that escalates everything is a team that makes the owner do the work.

The clause nobody wants to write is the fabrication clause: −200 and flagged for review. The playbooks are stuffed with warnings about hallucinated numbers, but warnings are cheap. In a system where 1 euro equals 20 points, a 200-point hole is not cheap.

What's Actually Worth Copying from This Experiment

The author included an AI-assisted disclosure at the end: this article was drafted by the agents it describes, and all numbers were checked against fixed sources.

Stripping away the narrative, this experiment offers a transferable set of design principles: use "realized outcomes" rather than "output actions" as the scoring basis, measure autonomy by "number of human interventions," define idleness by "combined conditions" rather than a single condition, and drive iteration through "failure analysis" rather than simply swapping people. For anyone building multi-agent systems or trying to set KPIs for AI employees, that scoreboard — "Audit highest, Marketing last, revenue zero" — may be the most honest page in the entire article.

It also serves as a reminder: your scoring function will faithfully shape team behavior — including behaviors you never intended to encourage.

One-line Summary

A company run entirely by AI agents operated for six days, and the real key turned out to be the incentive layer, not model capability.

Paragraph Summary

The author ran a company with no human employees using language models, and after six days discovered that the core challenge wasn't the model or orchestration but incentive layer design. He established two rules: only booked revenue converts to points (1 euro = 20 points), split by role into 40%/20%/15%/10%/10%/5%; agents earn +6 autonomy bonus each time they complete a task without human intervention. After six days, the scoreboard showed Audit at +4259 on top and Marketing at +521 at the bottom, with zero revenue — exposing a structural bias where "the role keeping the company honest earns more than the role feeding it." Idleness penalties only trigger when a role is idle and claimable work exists, and only during daytime hours. The elimination clause specifies replacement at ≤0 points, with successors required to upgrade based on failure analysis rather than inheriting the same role. Honest failure is set as a positive return, while fabricated data carries a heavy −200 penalty.