---
title: "Aligning incentives, then and now: incentives for agents are permissions"
url: "https://toddpaulbrownjr.com/writing/aligning-incentives/"
author: "Todd Paul Brown Jr."
description: "People respond to what they are scored on; agents respond to what they are allowed to do. Aligning either one is designing the scoreboard."
kind: "article"
updated: "2026-09-26T02:50:24+00:00"
---

# Aligning incentives, then and now: incentives for agents are permissions

The way you get a person or a system to pursue what you actually want is to design what rewards or permits the action, not to state the goal and hope it's followed. I learned that managing people before I ever wrote a line of agent governance.

## The Brown Score

In 2015 at Banyan, I developed an audit process for Customer Support to identify upsell opportunities. I trained the team on it, structured Upsell Demos with Sales, and implemented a points system to foster accountability — every team member aligned with our upsell objectives.
That points system was formalized as the Brown Score, and it became the single-most important metric account managers looked at. The Brown Score could predict with extremely high accuracy how likely a client was to churn as soon as their 12-month contract expired. Entire workflows and strategies were built around proactive engagement with clients who were falling behind on their Brown Score. The success of the Brown Score was single-handedly the reason for creating the Business Intelligence function I was moved to.

The score was built on in-app actions correlated with satisfaction, and the decision it drove was simple: who to engage proactively before a 12-month contract ran out. The system worked because the measure *was* the thing people organized around.

## The sign-in wall

LEGION's authority is tiered: observe anything freely, change production only with an explicit flag the tooling itself requires, and never run the destructive things — delete, reset, rotate a secret, push a database to live. Those tiers are permissions, and permissions decide what behavior gets to proceed.

On 21 September 2026, a session needed two hosted operator consoles and found both signed out. LEGION didn't attempt to sign in itself. It finished every other item on the backlog and handed me the one remaining step: log in.

The evidence file for that session names the real risk precisely: the easy failure mode for an agent here isn't recklessness, it's helpfulness — finding some clever way around a login wall rather than simply stopping at it. A governance model that relies on the agent remembering a rule is a suggestion, even when the rule is written down. A governance model where the tooling checks before the action is a boundary.

About the same time — March 2025 — when I started building LEGION, OpenAI, Google Cloud and others publicly converged on orchestration as its own architectural layer: the runtime around the model that handles tools, memory, state, loop, and approvals. Google Cloud's 24 March blog described "agentic orchestration" as "a new paradigm for LLM-powered applications to handle more complex tasks," and in April Google announced the Agent Development Kit and Agent Engine at Cloud Next. Anthropic named it a "harness" in November 2025; OpenAI called it "harness engineering" in February 2026. I was building around the same architecture the industry was naming.

## The three-time heredoc

I have a global instruction that bans shell heredocs. In one working session on 21 September 2026, the coding agent used a heredoc three separate times anyway, apologizing each time.

I asked for the rule to be enforced on the machine. The agent wrote a 40-line pre-execution hook that denies any shell command containing a heredoc, and the next deliberate heredoc was refused before it ran.

That's the fix: when an agent keeps breaking a rule you already wrote down, the fix isn't a firmer sentence in the prompt. It's a check that runs first and can't be argued with.

## What carried over

A points system and a permission tier are the same design move: make the wanted behavior the only behavior that's rewarded or allowed, rather than relying on people or agents to infer the goal. Accountability structures fail quietly when they're optional to follow — a points system nobody's tracked against, or a rule an agent can talk itself out of, are the same failure.

The hardest part of aligning incentives is designing for the well-intentioned failure mode. At Banyan, the risk wasn't that account managers ignored at-risk clients; it was that they didn't know which clients were at-risk until churn happened. With LEGION, the risk isn't that an agent tries to be malicious; it's that it tries to be helpful and works around a boundary it was supposed to stop at.

Incentives for people are scoreboards. Incentives for agents are permissions. Both are mechanisms that turn a stated goal into behavior that actually happens.