Building an Agent Culture
The teams that ship agents you can trust don't have a secret model — they have habits that keep an acting system honest: measured, traced, gated, and owned. Your job is to fund and protect those habits, and to reward the right things.
Culture here is concrete. Quality is a shared completion number over a task suite, reviewed regularly — not a demo. Traces are standard, so any failure is explainable. Gates on irreversible actions are respected, not bypassed for speed. Real failures become new test tasks instead of being waved away. And someone owns the agent's reliability as their job, rather than it being everyone's and no one's.
As a leader you set the incentives, and with acting agents the stakes are higher than with chatbots. If you reward flashy autonomy demos, you get reckless agents. If you consistently ask "what's completion over the suite, show me a trace, and where are the gates?", the team builds a system that's safe to trust with real actions. Celebrate the boring dashboard that proves reliability, not the demo that dazzles for sixty seconds.
Speed vs. the gate
Under deadline pressure, an engineer suggests removing the human approval on a risky action "just to hit the launch date — we'll add it back later." How do you respond, and what does your answer signal?
Hold the gate. Removing the human approval on an irreversible action to hit a date trades a schedule win for the risk of an unrecoverable, possibly public mistake — exactly the asymmetry that makes acting agents dangerous. Reframe it: "the gate is why we're allowed to ship an agent that moves money at all." What you permit under pressure becomes the culture; bypass a safety control once for speed and you've taught the team that gates are optional. Slip the date or reduce the agent's autonomy instead — never its safety.
The team that made the trace a ritual
A product team kept shipping agent changes that fixed one behavior and quietly broke another, because reliability lived in scattered heads and nobody routinely looked at what the agent actually did. Their manager made one change: a shared dashboard — completion, steps, escalations — plus a habit of reading a couple of real traces together at the start of each week.
Within a quarter the "fixed A, broke B" pattern faded, because no change could move the shared completion number or introduce a weird trajectory without someone noticing. Reading traces also surfaced risky shortcuts the agent was taking that no single metric had flagged. Nothing about the model changed; the manager had made the agent's behavior visible and collective — and five minutes a day of shared attention did more for reliability than any upgrade they'd tried.
Tessera reviews the agent's completion dashboard and a sample trace at the start of each weekly ops meeting. Reliability stops being the eng lead's private worry and becomes a shared number the team owns — and the gates hold, because bypassing one would be visible to everyone.
Quiz · Chapter 11
- A healthy agent culture treats reliability as:
- Real agent failures should become:
- Bypassing a safety gate "to move faster":
- What a leader rewards: