Deploy one agent on one workflow that is high-frequency, low-risk and currently getting dropped: follow-up is usually the right first choice. Give it real access to your tools rather than a description of them, set explicit boundaries on what it may do unsupervised, run it for two weeks against a baseline you measured beforehand, and only then add a second agent.
Most first agent deployments fail in one of three ways: the workflow chosen was too rare to produce a signal, the agent was given a description of the business rather than access to it, or nobody defined what it was allowed to do alone. All three are avoidable in an afternoon of planning.
Step 1: Pick the workflow, carefully
The right first workflow has four properties.
- High frequency. Daily at minimum. A weekly workflow takes a quarter to tell you anything.
- Currently dropped. If it already happens reliably, automating it saves less than you think.
- Low blast radius. A bad follow-up email is recoverable. A bad invoice is not.
- Measurable. You need a number that moves.
For most startups this is lead follow-up. For some it is inbound qualification or onboarding sequences. It is almost never content generation, despite being the most popular first choice: content volume is easy to increase and hard to attribute.
Step 2: Measure the baseline first
Before anything is deployed, write down the current numbers. How many leads got a second touch last month? How many hours went into it? What was the response rate? This takes twenty minutes and is the difference between knowing whether the agent worked and having an opinion about it.
Step 3: Give it access, not a description
This is the step that separates a useful agent from an expensive chatbot. An agent that has been told about your CRM will write plausible things. An agent connected to your CRM will write true things. Connect the real systems: inbox, calendar, CRM, docs.
The practical minimum for a sales agent is your CRM, your calendar and your email. For a marketing agent, your analytics, your publishing surfaces and wherever your brand voice lives.
Step 4: Give it the context a new hire would need
The useful mental model is a competent contractor starting Monday. What would they need on day one?
- Who you sell to, specifically, and who you do not.
- What you actually do, in the words customers use rather than your own.
- The three objections you hear most, and how you answer them.
- What a good outcome looks like: a booked call, a reply, a qualified lead.
- The tone. Paste in three emails you were happy with.
Twenty minutes here is worth more than any amount of prompt engineering afterwards.
Step 5: Set explicit boundaries
Write down, in plain language, three lists: what the agent may do without asking, what it must draft for approval, and what it must never touch. Most trust failures come from a founder discovering an agent did something reasonable that they nonetheless did not want done.
| Action | Autonomy |
|---|---|
| Qualify a lead against the ICP | Unsupervised |
| Update CRM records | Unsupervised |
| Send a follow-up to a warm lead | Unsupervised after week one |
| First outreach to a named target account | Draft for approval |
| Anything mentioning price or terms | Draft for approval |
| Contacting an existing customer | Never, initially |
Loosen these deliberately as the agent earns it. Tightening them after an incident is much more expensive than starting conservative.
Step 6: Run it for two weeks, and read the output
Actually read what it sends, every day, for the first week. Not to approve each item, to learn where its context is thin. Every awkward email is a missing piece of business context, and fixing that is a two-minute edit rather than a platform problem.
Step 7: Decide, then add the second agent
At two weeks, compare against the baseline from step two. Three outcomes:
- Clearly better: expand scope on this workflow before adding a new agent.
- Roughly neutral: almost always a context problem, not a capability one. Go back to step four.
- Worse: check whether the workflow was genuinely high-frequency and whether the agent had live access. It is nearly always one of those two.
Only then add a second agent. The compounding value in an agentic system comes from agents sharing context, the marketing agent knowing what sales heard, and that only works if the first one is genuinely well-briefed. More on how that coordination works in multi-agent orchestration, explained.
The three failure modes again
If a deployment goes wrong, it is almost certainly one of: a workflow too rare to measure, an agent describing your business rather than connected to it, or undefined boundaries. Check those three before concluding agents do not work for your case.
Key takeaways
- Pick a workflow that runs at least daily. Weekly workflows take months to produce a signal.
- Context beats configuration: an agent with live tool access and a clear ICP outperforms one with an elaborate prompt.
- Write down what the agent may do without asking. Ambiguity here is where trust breaks.
- Measure the baseline before you deploy, or you will never know whether it worked.
- Two weeks is the right evaluation window: long enough for signal, short enough to matter.