The agent retention problem: which agents survive year 1

Most AI agent platform marketing focuses on agent CREATION. Templates! Visual builders! Easy setup! Get your agent running in minutes!

What gets less attention: agent RETENTION. Of the agents you create, how many are still running and delivering value 12 months later?

This is the most important metric in agent platform success that nobody measures publicly. We've been tracking it internally. Here's what we've learned.

The data

Across our deployed customers, we've been tracking agent retention. At the 12-month mark for a cohort of agents created in 2024-2025:

→ 100% created → ~67% still running at month 3 → ~52% still running at month 6 → ~41% still running at month 12

The decline is real and consistent. About 60% of agents created get retired within a year.

Why? Multiple reasons. The interesting question is which agents survive and which don't.

Why agents get retired

When we investigate retired agents, the reasons cluster:

Reason 1: Workflow changed.

The business process that the agent automated changed. The agent no longer fits.

Frequency: ~30% of retirements. Common in fast-growing companies where processes evolve quickly.

Reason 2: Better alternative emerged.

A more general agent replaced multiple narrow ones. Or a new template made the custom agent obsolete.

Frequency: ~20% of retirements.

Reason 3: Not enough value.

The agent worked but didn't save enough time to justify maintenance attention.

Frequency: ~20% of retirements.

Reason 4: Quality drift.

The agent's output quality degraded over time and the team didn't tune it. Eventually retired in frustration.

Frequency: ~15% of retirements.

Reason 5: Owner left.

The person who built the agent left the company. Nobody else maintained it.

Frequency: ~10% of retirements.

Reason 6: Connector broke.

External service changed API. Agent stopped working. Wasn't fixed.

Frequency: ~5% of retirements.

Which agents survive

Looking at the 41% of agents still running at month 12, common characteristics:

Survivor pattern 1: Solves recurring high-volume pain.

Agents that handle work happening daily/multiple-times-per-day, where manual handling is genuinely painful, tend to survive. Examples: Sophia (meeting follow-ups), Carlos (pipeline digest), Anna (news digest).

Survivor pattern 2: Has clear human owner.

Agents with one person who explicitly owns them survive. Agents that are "team owned" (= nobody owned) get neglected and retire.

Survivor pattern 3: Has been maintained.

Agents that get the maintenance attention described in [post 187] survive. Agents that ran without tuning eventually drift into uselessness.

Survivor pattern 4: Solves work that ALSO requires human judgment.

Counter-intuitive: agents that completely automate something tend to get forgotten. Agents that surface drafts/recommendations for human review stay relevant because humans interact with them regularly.

Survivor pattern 5: Integrates with daily workflows.

Agents whose outputs land in tools the team uses every day (email, Slack) survive. Agents that produce outputs in separate places get forgotten.

What this means for new agent deployment

If you want agents that survive past year 1:

Choose agents that solve recurring high-volume pain. Don't build agents for occasional work. The setup cost won't be justified.

Assign clear ownership. Every agent should have one human owner who's responsible for maintenance.

Schedule maintenance from day 1. Weekly review, monthly tuning, quarterly assessment. (See [post 187] for detailed schedule.)

Design for human interaction. Even autonomous agents should produce outputs humans see regularly. Outputs landing in inboxes/Slack reinforce ongoing relevance.

Integrate with daily workflows. Don't put agent outputs in obscure places. Put them where the team already looks.

What this means for our product

The 41% year-1 retention rate isn't great. It's better than we feared but worse than we'd like.

We've been working on improving it. Specific product changes:

Better default ownership. New agents now prompt for an owner during creation. Helps with the "team owned = nobody owned" problem.

Maintenance reminders. Avery NXR now nudges owners weekly with sample outputs to review. Helps with the "agents drift because nobody looks" problem.

Connector health notifications. When a connector starts failing, the relevant agent owner gets notified immediately. Helps with the "connector broke and nobody noticed" problem.

Template improvements based on retired agents. When agents in a category retire, we analyze why and improve our templates. Helps reduce future retirement rates.

Better retirement workflow. When an agent should be retired, we make it easy. Better than letting it linger and confuse things.

These changes are bringing the year-1 retention rate up. We'll have updated data later in 2026.

Why this metric matters

Agent retention is the metric that distinguishes platforms that produce durable value from platforms that produce one-time bursts.

A platform that helps you create 20 agents in a week is impressive. A platform where 16 of those agents are still saving you time 12 months later is more valuable.

Most agent platform vendors don't talk about retention because it's a hard metric. We're talking about it because we think transparency on hard metrics is the right approach.

What we'd tell other agent platform vendors

If you're building an agent platform, track retention. Specifically:

→ Of agents created in month X, how many are still running in month X+3, X+6, X+12? → Why do agents retire? Categorize the reasons. → Which agent types have higher retention? Build more of those. → Which deployment patterns have higher retention? Promote those.

Retention data informs product strategy in ways that creation data doesn't. The platforms that figure out retention will outlast platforms that just maximize creation.

What we'd tell buyers evaluating agent platforms

When evaluating, ask vendors: what's your agent retention rate at month 12?

Most won't have an answer. The ones that do, even with imperfect numbers, are thinking about the right metric.

Don't pick platforms based purely on how easy they make creation. Pick based on how likely the agents you create are to still be running and delivering value a year from now.

The framing

Creating agents is the wedding. Maintaining a relationship with deployed agents is the marriage.

Most platforms over-invest in the wedding (slick onboarding, impressive demos, fast setup). Most teams under-invest in the marriage (maintenance, ownership, evolution).

The platforms that mature will be the ones that help teams stay married to their agents, not just get them deployed.

Avery NXR is investing on the marriage side. Our maintenance tooling, owner assignment, audit ledger, and tuning capabilities are all about helping deployed agents stay valuable over time.

What this means for our customers

If you're an Avery NXR customer and your agents are surviving past year 1, you've figured something out. The patterns above probably apply.

If your agents are retiring at higher rates than 60%, something's off. Probably one of:

→ Wrong workflow selection (not high-volume recurring enough) → Unclear ownership (everybody and nobody owns the agent) → No maintenance schedule (agents drift) → Agents disconnected from daily workflow surfaces

The diagnostics are usually one of those four. The fix is usually small.

If you want help diagnosing your specific agent retention pattern, we can help. Our customer success team has seen many patterns.

The bigger lesson

Most product metrics focus on adoption. The interesting metric is whether what got adopted continues delivering value over time.

For AI agents specifically — given the rate of change in models, business processes, and team composition — the retention question matters a lot.

Teams and vendors that focus on retention will outlast teams and vendors that focus on flashy creation.

→ avery.software — Free Desktop tier. The platform optimizing for agents that survive year 1, not just agents that demo well.