We Hold Agents to a Higher Standard Than Ourselves. Good.

This week Ram M published a clear-eyed piece, How Employees Accidentally Cripple Organizations. Its argument is careful and humane. Companies are rarely brought down by a few bad people or one catastrophic decision. They are worn down by thousands of small, reasonable actions taken by good employees trying to do their jobs. The word that carries the whole piece is accidentally. Nobody intends the harm. It emerges from incentives and structure and the ordinary wish to avoid risk and protect a relationship.

He names five mechanisms: self-preservation, local optimization, information distortion, dependency creation, and risk avoidance. Against each one he sets a column he calls the Hidden Cost: the project slip that doubles before anyone admits it, the departmental win that becomes an enterprise loss, the decision leadership makes on an incomplete picture, the knowledge that walks out the door, the opportunity that passes while everyone waits for certainty. The harm is real and quantifiable. It is just never invoiced to anyone.

I read the list a second time and noticed something. Every one of those five behaviours is something we have already decided is unacceptable in an AI agent. We write requirements against them. We refuse to deploy systems that do them. We hold agents to a standard of conduct that the organisations deploying them have never been held to.

We hold agents to a higher standard than we hold ourselves.

This post is about why that is a good thing, and worth defending.

The higher bar, drift by drift

The five are not failures of character. They are what happens when you route decisions through people who are, quite reasonably, protecting a position, a metric, a relationship, or a reputation. The third one, information distortion, is the sharpest case: it is the bullwhip effect turned vertical, and it gets its own section below. So take each in turn, with no blame attached to anyone. On the left, the accidental drift and the hidden cost Ram names. On the right, the standard we hold an agent to instead.

Note what the right column is not. It is not “agents are honest and people are not.” Each standard is something we insisted on when we built the agent, and never insisted on when we built the organisation. The agent does not surface the bad news because it is virtuous. It surfaces it because it has no career to protect and no review meeting to survive, and because the Decision Trace it writes to is append-only. We engineered the absence of the pressure that makes a good person wait.

Two bullwhips, one human cause

We have known about the horizontal bullwhip since the 1960s. Jay Forrester described the dynamic in Industrial Dynamics in 1961, having sketched it in a 1958 Harvard Business Review article, and he built the MIT Beer Game to teach it. Peter Senge made that game famous in The Fifth Discipline; Lee, Padmanabhan and Whang gave the effect its name in 1997. But the part that matters most is the part the Beer Game was built to prove, and that John Sterman measured in 1989: the amplification is behavioural. The players are not short of data. They panic, they hoard, they over-order to protect themselves, they optimise their own link and misread the feedback from everyone else’s. The bullwhip is simply what a chain of people does under local incentives and partial visibility.

So we spent sixty years building process around it (S&OP, collaborative planning, vendor-managed inventory, demand-driven replenishment, control towers, shared forecasts) and we still have not cured it. Every one of those is a structural workaround for a human reflex, and the reflex is still there: we hold the line on a good week and watch the oscillation come back on a bad one.

Now read Ram M’s five mechanisms again. Self-preservation, local optimization, information distortion, risk avoidance: that is the same set of human reflexes that drives the horizontal bullwhip, simply pointed in a different direction. The signal that softens as it climbs the management hierarchy is the vertical bullwhip, the identical disease running up the org chart instead of up the supply chain. Same cause, a different axis. The only real difference is that we gave the horizontal one sixty years and a research literature, and the vertical one got a name in a LinkedIn post this week.

The asymmetry is almost comic when you hold the two side by side. A 2x amplification across four supply-chain echelons triggers a cross-functional task force; the identical 2x amplification across four management layers gets a fond name, “managing up”, and a shrug. We would never deploy an agent that quietly rewrote a number softer each time it passed it along: that is the fabrication failure we screen agents for before they go anywhere near production. We have tolerated exactly that behaviour in our own org charts for a century, because it wears a tie and means well.

Because the cause is human, the only real cure is one that takes the human reflex out of the loop, not one more process laid on top of it. That is what an agentic substrate does, on both axes at once. The agent has no position to protect, no local metric to panic over, and no relationship to keep sweet, so it does not amplify. Radical visibility gives it one shared world model and calibrated state that passes between tiers unchanged, horizontally across the echelons and vertically up the decision stack. And the goal is shared by design, not only the signal: when one agent’s decision would touch a parameter another owns, it must first play the change out in a parallel scenario (the network’s DAG and the other agents simulating it) and may only propose it if the candidate plan improves the enterprise scorecard rather than its own local cost. Sixty years of process never reached the cause. The substrate starts there.

Why the higher standard is a good thing

It would be easy to read all this as a complaint: that we are unfair to agents and ought to relax. That is exactly backwards. The higher bar is not a burden we have unfairly placed on agents. It is the right bar, for four reasons.

1. It is the only bar a built thing can actually meet. You cannot engineer self-preservation out of a person. You can ask them to rise above their own incentives, every day, unsupported, and the best people often will. But you are asking, not guaranteeing. You can engineer it out of an agent. The higher standard is not unfair to the agent. It is the standard that only becomes available once a thing is made of code instead of career. We hold a suspension bridge to a higher load than a rope bridge, and no one calls it prejudice against rope.

2. The higher standard is the entire reason to deploy them. If you held an agent only to the standard you hold yourselves, locally optimal, self-protective, the message softened on the way up, it would be no improvement on the organisation you already have. The whole case for the agent is that it clears a bar the organisation structurally cannot. Lower it to the human bar and you have spent the budget to reproduce the Hidden Cost column, only faster.

3. The high bar is the first yardstick the organisation has ever had. This is the one that matters most. The five drifts are tolerated because they are invisible. There is no reference standard to measure them against, which is exactly why the costs stay hidden. Stand an instrumented agent next to the organisation, every decision traced, every estimate carrying a calibrated likelihood, every choice scored against the enterprise, and for the first time people can see what undistorted flow and enterprise-optimal actually look like. Holding agents to the high bar does not leave humans at the low bar. It hands the organisation the measuring stick it never had, and the Hidden Cost column stops being hidden.

4. It is what earns the right to free the human. You can only take a person out of the bottleneck and hand a decision to an agent if you have held that agent to a standard worth trusting. The high bar is precisely what lets a person stop being the approval-relay and become the Override authority, the one role that genuinely needs a human. The standard we demand of the agent is what buys back the human’s time for judgement. Hold the agent low and you are back to a human checking every decision, which is the bottleneck Ram’s whole article is about.

Agents have their own accidents

None of this means agents are clean. They can miscalibrate. They can act outside the scope you intended. On the narration layer, a language model can say something that is simply not so. If I stopped at “agents do not have the five flaws,” I would be making a boast, not an argument.

The point is the opposite. We hold the agent’s own failure modes to the same engineered standard: a conformally calibrated band on every estimate instead of a false point of confidence, a guardrail that holds any decision crossing a bound, a standing human Override, and a hard rule that the language model is never on the decision axis. Same standard, applied to the agent’s accidents and the human ones alike. The difference is not honest versus dishonest. It is instrumented versus invisible. An agent’s accident leaves a calibrated, audited trace and trips an alarm. The five human accidents are, by their nature, the ones no one sees until the Hidden Cost has compounded into a crisis.

The standard was always the right one

Ram M closes by naming the work for leaders: design systems that reward transparency over self-protection, enterprise thinking over departmental optimization, truth over comfort, knowledge sharing over dependency, and informed action over excessive caution. Read that again as an engineer rather than a manager. It is a specification. It is, almost line for line, the specification we hold our agents to. He wrote it as the standard human organisations should meet and mostly cannot. We took the same standard and made it executable.

So yes, we ask more of our agents than we ask of ourselves. Keep doing it. The higher bar is the only one worth building to, the only one that justifies the build, and, once an agent is meeting it in plain sight, the first honest yardstick an organisation has ever had to measure its own accidental drift against. The Hidden Cost column was always there. We just never had anything that refused to pay it.

---

Further reading.

The piece this post responds to.

• Ram M (June 2026). How Employees Accidentally Cripple Organizations. LinkedIn: How Employees Accidentally Cripple Organizations. The five-mechanism framing, and the Hidden Cost column, this post takes as its starting point, and whose insistence on the word accidentally it tries to honour.

Azirella’s expression of the standard.

• AI·IO·ML: the operating model that lets an agent act by default, keeps the human’s override first-class, and measures every outcome to learn from it, so neither self-protection nor approval-paralysis is the resting state.
• The Decision Flow Problem: why decision velocity, not headcount, is the constraint, and why local optimization is the tax on it.
• Stop Using Averages: the conformal layer that puts a calibrated band on every estimate instead of a comfortable point.
• Operating Knowledge: the substrate that elicits tacit expertise into a shared, lifecycle-managed store rather than leaving it in one person’s head.
• How agents learn: the hash-chained Decision Trace and the override-to-training loop that keep the record honest and the system improving.