The vocabulary around AI agents has settled into three tiers — assistant, semi-autonomous, autonomous — and most teams pick the wrong one for the wrong reason. They either over-buy autonomy because the demo was impressive, or they under-buy and end up with a chat window that costs as much as a junior hire and produces less.
This post is the framework we use to choose, and the agents that actually deliver at each tier.
The three tiers, defined
Strip the marketing out and the tiers map to a single variable: who is responsible for the next step?
- Assistant. You ask, it answers, you decide what happens next. The agent never moves without you.
- Semi-autonomous. The agent plans and executes a multi-step workflow, but pauses for approval at high-stakes checkpoints. You're the brakes, not the engine.
- Autonomous. The agent runs end-to-end and reports back when finished. You're a reviewer of outcomes, not a participant in the work.
The difference is not how smart the underlying model is. All three tiers use the same models in 2026. The difference is the guardrails around what the agent is allowed to do without checking in.
Why the wrong tier kills the project
We've watched two failure modes show up repeatedly.
Failure mode 1: too much autonomy, too early. A team gives an autonomous sales agent a contact list and asks it to send "personalized outreach." Two weeks later they discover it's been sending boilerplate to enterprise prospects with the wrong company name. The agent did exactly what it was told. The team didn't realize the cost of being wrong scaled with the prestige of the recipient.
Failure mode 2: too little autonomy, too long. A team pays $200/seat/month for an assistant-tier coding tool and spends 40% of every session approving micro-edits. The agent could ship the same PR autonomously in 1/10 the time, but the team is afraid to let it. They're paying agent prices for copilot value.
The framework below cuts both failures off.
The three-question test
For any task you're considering handing to an agent, ask:
1. Is the outcome objectively verifiable?
Can a script, a test suite, or a checklist tell you within a minute whether the agent succeeded? "PR passes CI" is verifiable. "Email sounds professional" is not.
2. Is the cost of a single mistake under $1,000?
A wrong unit test costs minutes to revert. A wrong email to a $50k account costs the relationship. A wrong line item in a tax filing costs an audit. Calibrate to your stakes.
3. Are you culturally OK with the agent acting without asking?
This is the question most teams skip. Some teams are fine with an agent shipping a PR overnight. Others want to read every word before it leaves the building. Neither is wrong, but knowing where your org sits saves a quarter of churn.
Three yeses → autonomous is fine. Any no → stay at semi-auto or assistant.
The autonomy map by category
Here's how the framework plays out for the categories we cover.
Code
The verifiability is excellent (tests, CI, type checkers). Mistake cost is low for most tasks (revert the PR, lose an hour). Most teams are comfortable with the autonomy — they've been letting CI deploy for years.
→ Autonomous is the right default. Devin and Sweep are the leaders here. Cursor Agent fits if you want a semi-autonomous fallback inside the editor.
Research
Verifiability is mixed. "Found three relevant studies" is easy; "synthesized a defensible recommendation" is not. Mistake cost varies wildly — bad research that goes into a slide deck is cheap, bad research that goes into a regulatory filing is not.
→ Semi-autonomous wins. Manus and Perplexity Labs both default to this mode: they go deep, but pause to confirm the brief before producing the final artifact. Use Elicit for literature-heavy work where citation accuracy matters.
Support
Verifiability is high once you measure CSAT and deflection rate. Mistake cost is moderate (a bad answer doesn't break the company but compounds across thousands of tickets). Most teams have the operational maturity to let this run.
→ Autonomous. Sierra is the leader for branded customer-facing work; Decagon and Parloa are the alternatives. Keep a human in the loop for escalations, not for tier-1.
Ops / personal
Verifiability is low. "Triaged my inbox correctly" is judged by you, in your head, three days later. Mistake cost varies — sending a meeting decline to the wrong person is mild; auto-paying the wrong invoice is not.
→ Semi-autonomous. Lindy is the right shape: it builds workflows that pause at approval gates. Same for Relay Agents. Avoid autonomous mode for anything that touches money or external relationships until you've watched it for three months.
Sales / marketing
Verifiability is low (open rate doesn't tell you whether the email was on-brand). Mistake cost is moderate to high depending on recipient prestige. Most teams overestimate how comfortable they are letting an agent represent the brand.
→ Start at assistant, graduate to semi-autonomous. Artisan Ava and Clay both support semi-auto modes — use them. Don't run anything in full autonomous mode until you've audited 500 sends manually.
The hidden cost of choosing wrong
We track click-through and time-to-value across the agents we index. The pattern is consistent: teams that pick the right tier on day one ship value in the first week. Teams that pick wrong spend 3–6 weeks oscillating between "this is magic" and "this is unsafe," and most of them end up uninstalling.
The agents are good enough now that the choice between two well-designed options at the same tier rarely matters. The choice between tiers always does.
Recommendations
Three concrete picks for the most common situations:
- You're a working engineer who wants help shipping faster. Cursor Agent (semi-auto), plus Devin for overnight backlog if you have the budget.
- You're a knowledge worker who wants to outsource the research stack. Perplexity Labs as your daily driver, Manus when you need a finished artifact instead of a brief.
- You're an ops person who wants to stop running a meeting-scheduling and inbox-triage loop yourself. Lindy at semi-autonomous mode. Take six weeks before you let it run unsupervised on anything that touches calendars or money.
The framework will outlive any specific recommendation. The agents change quarterly. The question "is the outcome verifiable, is the mistake cheap, am I culturally OK with autonomy" does not.