Prompt Optimization Is Here: Your Agent Now Improves from Your Own Call Data

Prompt optimization is live. Starting today, your agent can tell you where it's falling short — and propose exactly how to fix it.

The system prompt your agent runs on was written once. Maybe Workforce Wave generated it from your website. Maybe you wrote it yourself. Either way, it was based on what you expected callers to ask — not what callers actually asked once the agent went live.

Those two things are never identical.

The Problem with Set-and-Forget Prompts

Here's a pattern we see constantly in call data. A business provisions an agent. The agent performs well on the expected cases — appointment scheduling, hours, location, basic FAQs. Then real callers start calling.

And real callers ask about the after-hours emergency line. They ask whether the practice accepts a specific insurance they just switched to. They ask what happens if they need to cancel same-day. They ask about parking.

None of those were in the original system prompt. The agent handles them awkwardly — hedges, says it doesn't know, transfers when it shouldn't have to. Callers hang up. The business doesn't know it's happening because there's no report that says "your agent is silently failing on these five topics."

Until now.

How Optimization Works

Every week, for agents with 50 or more calls in the trailing seven days, Workforce Wave runs an optimization analysis:

Pattern identification — Workforce Wave reads the last 100 transcripts and looks for systematic patterns. Questions that didn't get good answers. Topics where the agent hedged or said "I don't know" more than twice. Calls that escalated to a human when they probably shouldn't have. Booking flows that were abandoned mid-conversation.

Evidence collection — For each identified pattern, Workforce Wave pulls the specific call excerpts that demonstrate it. Not just a count — the actual transcript snippets that show what went wrong and why.

Prompt revision — Workforce Wave generates a revised system prompt that addresses the identified gaps. New knowledge sections for underserved topics. Clearer instructions for the scenarios where the agent was inconsistent. Adjusted escalation logic where the current rules were too aggressive or too permissive.

Impact estimation — Workforce Wave projects the expected improvement based on how frequently the addressed patterns appeared in the transcript sample.

The result lands in your review queue as a prompt_optimization item.

What the Proposal Looks Like

The review queue shows the optimization proposal as a structured diff. You see the current prompt alongside the proposed revision, with the specific changes highlighted. Below the diff, you see the supporting evidence: the call excerpts that justify each change, with timestamps and call IDs so you can listen to the original recording if you want.

The projected impact estimate sits at the top: "Based on 23 calls with this pattern, this change is projected to reduce escalations by approximately 18%." That number isn't marketing — it's a direct calculation from your own call data.

You can approve the full proposal in one click, reject it, or approve individual sections. If you approve a section but want to rewrite the specific text, the prompt editor is one layer down.

For API Users

Optimization proposals appear in the review queue with type: reviewitem.promptoptimization. Query them directly:

curl "https://api.workforcewave.com/v2/review-queue?type=prompt_optimization" \
  -H "Authorization: Bearer $TOKEN"

To apply an approved optimization programmatically:

curl -X POST https://api.workforcewave.com/v2/agents/{id}/prompt \
  -H "Authorization: Bearer $TOKEN" \
  -d '{ "optimization_id": "opt_abc123", "approved": true }'

The agent's system prompt updates within minutes of approval.

What Doesn't Change

Your agent's core identity, persona, and knowledge base stay intact unless the optimization explicitly proposes changes to them. Workforce Wave is conservative about identity-level changes — it focuses on knowledge gaps and instruction clarity, not personality rewrites.

If you've customized your prompt significantly after provisioning, Workforce Wave's diff shows your version as the baseline. It's not comparing against some stored original — it's comparing against what your agent is actually running today.

---

Prompt optimization runs automatically every week for qualifying agents. If your agent has hit the 50-call threshold, the first proposal may already be in your review queue.

Go check. Your call data has been telling you something.