AI agent skills on a CV: what to write in 2026

What does the "self-driving company" mean for your CV?

Your CV now has to prove you can direct and check an agent's work, because the work itself is no longer the scarce part. Speaking at a San Francisco tech conference this week, Replit co-founder and CEO Amjad Masad described his own company as a "self-driving company" — self-improving, with agents handling much of the grunt work while people get valued for judgment and taste. His example was support triage: an agent reads the incoming ticket, the GitHub issues, the codebase and the Linear tickets, then hands a human the full context and a decision to make. Shipping the patch is a one-click handoff to another agent. Notice what's left for the person — the call, not the keystrokes.

Masad was clear that the human stays in the middle of the loop. "You need human judgment; you need taste," he said, arguing agents change what people work on rather than removing them: more time deciding what to build, more whiteboarding, more architecture. He also said every job changed over the past year and that much of America and the world still doesn't know the change is coming. That gap is your opening. Most CVs landing on a recruiter's desk this month still read like a 2019 job description — a tidy inventory of duties performed by hand, with no sign of who was supervising what.

Here's the uncomfortable part. If an agent can do the task you spent three bullets describing, those bullets now argue against you. A hiring manager reading "processed 200 invoices weekly" in late 2026 doesn't see diligence; they see a cost line waiting to be automated. The same experience written as "owned the invoice agent's exception queue, caught 1 in 12 mismatches before payment" reads as judgment. Same job, same person, completely different value. Start by sorting your current bullets into two piles: things a competent agent could do unsupervised, and things that required you to decide, verify, escalate or refuse. The second pile is your CV.

How do you list AI agent skills on a CV without sounding fake?

Put agent skills inside your experience bullets where the work actually happened, and keep the tool names to one short line near your skills. A standalone "AI skills" block listing six products proves you have subscriptions, not competence. The formula that survives scrutiny: verb + the agent or workflow + the scope + the oversight mechanism + the outcome with a number. For example: "Ran a three-agent research-to-draft pipeline for 40+ client briefs a month; built a 12-point review checklist that caught factual errors before publication and cut rewrite cycles from three to one." That bullet names no vendor and still makes the skill unmistakable.

Specifics beat adjectives every time. "Leveraged cutting-edge agentic AI to drive efficiency" tells a reader nothing and trips the same alarm as any other stock phrase. Instead, say which step the agent owned, which step you owned, and where the handoff sat. Three details do most of the heavy lifting: the trigger that sent work to a human, the thing you looked at to decide, and what happened when the agent got it wrong. If you can't name the escalation rule you worked to, you probably weren't supervising — you were just prompting, and that's a smaller claim. Make the smaller claim honestly rather than inflating it.

Order matters as much as wording. Lead each role with the bullet that shows the widest scope of judgment, not the one with the shiniest tool in it. Juniors should put agent oversight in the role where it happened and treat university or side projects as supporting proof; seniors should fold it into scope language — how many workflows, how much spend, which decisions were yours to sign off. If drafting from a blank page stalls you, a chat-based CV builder will pull the detail out of you role by role, and the method behind building a CV by talking to an AI assistant works the same way manually.

Which numbers prove you can supervise AI agents?

Four numbers carry the most weight: how many agents or workflows you supervised, the error or defect rate you caught and kept out of production, hours or cycle time returned to the team, and cost per task or per run. Those four map directly onto what agent teams actually track — task success rate, tool-call error rate, latency and cost. Quote them as a before-and-after pair wherever you can: "cut manual tagging from eight hours a week to twenty minutes" lands harder than "improved efficiency by 95%", because the baseline tells a reader you measured something real rather than reverse-engineering a percentage.

Get the baseline from records, not memory. Time comes from process logs, ticket timestamps and sprint data; error rates come from QA reports, defect logs and rework tickets; volume comes from transaction or ticket counts; cost comes from invoices and platform billing. If you never captured a baseline, count the volume for one representative week and multiply, then label it honestly as an approximation — "roughly 30%" is defensible, a suspiciously precise 31.4% is not. Interviewers will ask how you arrived at the figure, and "I pulled the queue export for Q2" is the answer that keeps the conversation going.

Two more metrics quietly separate the people who supervised agents from the people who watched them run. Override frequency — how often you rejected or corrected an agent's output — proves you were reading the work, not rubber-stamping it. Review throughput proves you didn't become the bottleneck, which is the classic failure mode when checkpoints are set too tightly and the human turns into the rate limiter. If your record-keeping is thin, the discipline of keeping a dated log of what you shipped pays off; the same logic applies to turning dated work records into quantified bullets rather than trusting recall at CV-writing time.

What job titles and keywords should you use for agent work?

Use the title the job ad uses, and mirror its exact phrasing in your CV. The oversight role has crystallised under a handful of names now carrying real salary bands: AI agent manager, agent supervisor, AI ops manager, agent QA lead and AI orchestration specialist. They aren't interchangeable. Agent supervisor work centres on monitoring a portfolio of deployed agents, spotting unexpected behaviour and triaging failures. AI ops leans toward uptime, evaluation frameworks and cost control. Agent QA lead is test cases, golden datasets and regression rates. Pick the one that matches what you did, and if your internal title was something vague, put the market-recognised equivalent in brackets after it.

On keywords, parsers still match literal strings, so the phrases themselves matter. The vocabulary hiring teams screen for includes human-in-the-loop review, evaluation design, guardrails, escalation paths, agent observability, prompt and context engineering, tool-call reliability and cost per task. Governance language has also moved from nice-to-have to expected, partly because the EU's AI Act brought high-risk obligations into force in August 2026 — human oversight, audit trails and traceable agent actions. Even outside Europe, employers now ask who approved a run and when. Write those words where they belong, in the bullets that describe the work, not stacked into a keyword block at the bottom.

If your employer hasn't deployed agents yet, build the evidence yourself and label it accurately. A side project with a documented evaluation set, a failure log and a short write-up of what you changed after the agent got something wrong is stronger proof of oversight than a certificate. Put it under Projects with the same structure as an experience bullet: scope, check, result. Juniors get more mileage from this than they expect — the rules for putting an AI-built project on your CV apply directly here, especially the part about being ready to explain every design decision in the interview.

What should you cut from your CV now that agents do the tasks?

Cut the duty lists, the tool inventories and the AI adjectives. Any bullet that starts "responsible for" and describes a repeatable, rule-following task is now arguing that your job can be automated, so either rewrite it around the judgment you applied or delete it. The same goes for the 2024-style stack bragging — a row of framework and vector-database names with no shipped outcome attached reads as table stakes at best. And "AI-powered", "cutting-edge" and "transformative" should go on sight; they're padding that makes a reader assume the sentence has nothing underneath it. A free CV check that flags AI-sounding phrasing will catch the ones you've stopped noticing.

Cut length, too. Three bullets about a process an agent now runs should collapse into one about the exception handling you owned, which buys you room for the decisions, trade-offs and escalations that actually differentiate you. Watch for the quieter problems while you're in there: generic summaries, bullets with no numbers anywhere, and layouts that break parsing. Those are still the most common reasons strong candidates get filtered, and the pattern behind the bullets and layouts that get CVs rejected hasn't changed just because the work has. A two-column template full of icons will fail before anyone reads your oversight story.

Here's my contrarian take: writing "AI" into every bullet is now a weaker move than writing it into two. When every applicant claims agent fluency, the differentiator is the one place you say no — the project where you decided an agent wasn't appropriate, or the output you rejected and why. Keep that bullet. Then test the result rather than guessing: a free CV analysis scores impact and ownership alongside ATS compatibility and clarity, which is precisely the axis this shift is measured on. Paste the job ad in too, and you'll see which agent-oversight keywords the role wants that your CV is still missing.

Frequently asked questions