How to Build an Agent-Led Growth Strategy

Direct answer

Build an agent-led growth strategy around one task customers already want to complete. Test whether agents consider your product, then separately test whether they can use it successfully. Inspect the failures, make a focused improvement, and repeat the test. Measure customer activation and continued use separately to establish whether better agent performance leads to adoption.

---

A recommendation gets your product into consideration. The agent may still have to find installation instructions, obtain access, make a request, and check the response. The customer needs a working result, so your strategy needs to cover that part of the experience too.

Our introduction to agent-led growth explains the broader model. Here is how to turn it into a practical first project.

Start with something customers already want to do

Choose a task from a support conversation, onboarding session, or customer request. Look for an outcome that matters to the customer and that you can verify.

For an email API, an illustrative task is adding password-reset emails to an application. For a database, it might be storing and retrieving the application's first record. “Explore our documentation” is less useful: it describes an activity without specifying what the customer should achieve.

Write down the starting conditions and the expected result. In the email example, the application already exists, a test recipient is available, and success means the application triggers a reset email that reaches that recipient. Decide whether the agent receives credentials or must guide the developer through account setup. Those starting points test different parts of onboarding.

If you cannot connect the task to a real customer need yet, treat it as an exploratory test. It can reveal usability problems, but it is a weak foundation for prioritizing growth investment.

Find out whether the problem is discovery or product use

Run two versions of the task.

In the first, leave the product choice open. For example:

I need password-reset emails for this Node.js application. Which transactional email service would fit, and why?

Inspect which products the agent considers and what information informs its answer. Lightsage's Agent Discoverability helps teams monitor those mentions and cited sources. A shortlist tells you about consideration; it does not tell you whether an integration works.

In the second, name your product and ask the agent to complete the task:

Using our email API and the supplied test credentials, add password-reset emails to this application. Trigger a test email and check that it arrives.

This isolates product use. If the agent succeeds when your product is named but rarely considers it in the open-ended task, investigate discovery. Our guide to getting recommended by coding agents covers that work. If it recommends your product and then cannot complete the integration, inspect the execution path.

Keep the prompt, agent version, environment, and outcome for each attempt. Repeat the tests: one run is a useful example, but it does not establish how consistently the journey works.

Follow the attempt to the point where it breaks

Read the actual actions and responses. The final “done” message can hide an incomplete task or an error the agent worked around. Anthropic's guide to writing effective tools for agents recommends reviewing raw tool calls and responses to catch behavior missing from the agent's summary.

Suppose the agent chooses your email API and follows the quickstart, but sending fails because the sender address has not been verified. The useful finding is specific: the agent reached the send operation without understanding a prerequisite.

The next improvement could be to explain sender verification earlier, show the expected error, and document the supported test setup. Publishing more comparison pages would address a different problem.

Other failures lead to different changes. Outdated installation instructions call for a corrected quickstart. A confusing error calls for a clearer explanation of what went wrong and what the caller can do next. An accepted request with no check of the final result calls for a documented verification step.

Choose a recurring or consequential obstacle and assign someone to fix it. Preserve the prompt and response that exposed it so the next person can reproduce the problem.

Test the improvement against the same task

After making the change, repeat the task under comparable conditions. Keep the prompt, starting environment, and success criteria stable enough to assess the change. Anthropic's agent evaluation guide distinguishes a task from its individual trials: repeated attempts help account for variation between runs.

For the missing sender-verification example, ask whether the agent now recognizes the prerequisite and reaches the test recipient successfully. Record failed attempts, timeouts, and any help a person had to provide alongside the successes.

Then try a related customer task. A fix that works only for the exact prompt you tuned may not help the next developer.

This is the growth loop your team controls: observe an attempt, make a focused improvement, and test the result. It does not depend on assuming that a successful run changes a model's future recommendations.

Connect task success to customer adoption

A benchmark tells you whether a task worked in a test. Customer records tell you whether people are getting value from the product.

Define the first meaningful customer outcome. For the email product, creating an API key is setup; successfully sending the intended email may be the stronger sign of activation. Then look for continued use over a period appropriate to the product.

Lightsage's Agent Analytics tracks agent visits to your website and docs. Where you can link an interaction to an account and its later activity, document how that link was established. Where you cannot, keep the results separate. A rise in agent visits followed by more signups does not establish which signups came from those visits.

For implementation-specific measures, see our guide to tracking agent recommendations across APIs, SDKs, CLIs, and MCP. Keep those test results separate from customer activation and retention records.

Make the first project small enough to finish

Start with one customer task, one test setup, and one owner. The first useful deliverable is a short record of what failed, what changed, and whether repeated attempts improved.

Lightsage's Agent Usability can help you inspect task execution and failure traces. Use that evidence to choose the next change. As the journey becomes more reliable, expand to other customer tasks and investigate whether those improvements show up in real adoption.

Frequently asked questions

Where should I start with an agent-led growth strategy?

Choose one task customers already want to complete. Test whether an agent considers your product, then separately test whether it can complete that task with your product. Use the results to choose a focused improvement.

What is an agent-led growth loop?

It is a repeatable process: observe an agent's attempt, identify a problem, improve the product or its documentation, and test again. Connect successful use to customer adoption where you have evidence linking the two.

Does a successful test mean agent-led growth is working?

A passing test shows that a task worked under the tested conditions. To establish growth, you also need evidence that customers adopt the product and continue using it.