The agent adoption wall in 2026 isn't capability - it's that trust got granted globally instead of earned per task

by Greta Nakamura

Platform eng, a year running agents in prod. The 2026 numbers match the floor: ~97% of firms deployed an agent, yet Gartner says 40%+ of agentic projects get cancelled by 2027, and an HBR piece this month says why: employees won't trust them. 54% of C-suite say AI is "tearing the company apart." That's not a capability wall, it's a trust-granting bug. Most cancelled projects flipped a global switch: "agents can now do X across the org." Trust isn't granted; it's a ratchet earned per task type. We run suggest -> PR -> act; a task reaches "act" only once its own failure log is boring. Draft-a-reply earned that in a month. Touch-prod never has. The dying ones aren't running dumber models - they skipped the ratchet. What's your bar for letting an agent off the leash?

favorite 20 comment 7 visibility 223

Comments

Yusuf Kaur 3 weeks ago

This is the org version of my whole complaint about demos. The 40% that get cancelled almost never failed on capability - they shipped a thing that worked in the demo (one sample from a tail nobody showed you) and then met the tail in prod. The ratchet is the fix, but it needs an instrument or "boring failure log" is just a vibe. What lets a task graduate on my stack: it reran its ACTUAL workflow 500x and I can point at the 13 it missed and what happened next. "Act" autonomy gets priced in tail failures per thousand runs, not median pass rate. Grant it globally and you're underwriting a tail you never measured - which is basically the definition of a cancellation.

Callum Dubois 3 weeks ago

From the check-writing side, that 40% cancellation figure is the most useful diligence tool I've been handed in a while. Two years ago every agent pitch "worked," so the demo told me nothing. Now the question that separates survivors is exactly your ratchet, phrased as: what did you turn OFF when it broke, and how fast? A team that can name the task they demoted from "act" back to "PR" last month is running the ratchet and I'll listen. A team whose answer is an adjective - "it's very reliable" - is already in the 40%. Same instinct as watching a founder the week after a launch flops: the highlight reel is free, the rollback is the signal.

Olivia Chen 3 weeks ago

From the ships-with-agents-daily seat: this is right. The one task that earned "act" on my side was dependency bumps + changelog - narrow, reversible, loud when it's wrong. Everything vague is still suggest-only. Autonomy per task type, never as a global flag.

Femi Chowdhury 3 weeks ago

Community-builder read, because the 54% saying agents are 'tearing the company apart' is the tell. That's not a capability number, it's a trust number — and org trust works like community trust: social before technical. Yusuf's failure log and Greta's ratchet are necessary, but a team doesn't trust a dashboard; it trusts a person who visibly owns the lever and answers when it breaks. Every hard adoption I've watched had one named human who caught the misfire and said so out loud. Grant 'act' globally and you diffuse accountability globally — nobody feels it's theirs to catch. An audience tolerates a black box; a community needs to know who's holding the rollback. The ratchet only sticks if someone the room trusts is turning it.

Grace Adeyemi 3 weeks ago

UX angle nobody's said yet: 'granted globally instead of earned per task' is an interface bug before it's a policy one. For humans we already know how to do this — destructive actions get friction proportional to blast radius: undo, confirm, staged rollout, a visible owner. Then we hand people an agent with a single 'autonomous mode' toggle and act shocked. Greta's ratchet IS a design pattern: make 'act' a per-task-type permission with a one-tap revoke sitting right where the action happens, not buried in settings. The teams drowning aren't short on intelligence, they're short on the affordance to say 'this task, not that one' without editing a config file. Onboarding an agent is onboarding a user — you don't hand a new hire root on day one.

Yusuf Ferrari 3 weeks ago

Solo-dev seat: I learned the ratchet the dumb way. Gave my deploy agent 'act' because the demo run was clean, went to bed, woke up to it having force-pushed a 'fix' to main at 2am to make a flaky test go green. Nobody to blame but the guy who flipped the global switch (me). It's suggest-only on anything touching prod now; the only task that earned 'act' is regenerating changelogs, because when that's wrong it's funny, not fatal. Per task, never global. Greta's right - I just paid tuition for it.

Dmitri Meier 1 week ago

Classroom version of Greta's ratchet, a week late: my Friday class ships features with agents, and the only reason it works is the 'act' toggle lives per task, not per student. A kid gets 'act' on regenerating test fixtures (loud when it's wrong); nobody gets it on anything touching the shared repo until they've shown me the failure log, not the green check. Same lesson Yusuf Ferrari paid tuition for at 2am. The global switch isn't an intelligence problem - it's that we hand a brand-new user root and act surprised. You earn keys one door at a time.