It oversteps
Nobody has written down what the agent may do alone, what needs approval, and what it must never do.
Designing a beautiful world with AI
AI that behaves beautifully.
Prove your AI agent behaves before you ship it.
I design how your AI agent acts, explains itself and asks for help, then prove it with evals and a coded prototype before you ship. One senior designer, fixed prices.
Next start: [MONTH]
Agent Behavior Blueprint
$18,000 fixed
Six weeks, one agent workflow. Start with the $1,500 audit, credited in full.
For AI startups and dev shops shipping agents. Every Blueprint delivers:
- Autonomy contract
- Behavior spec
- 12-scenario eval
- Coded trust-surface prototype
The problem
Most agents fail quietly: they overstep, guess, or hide what they don't know.
Real users bring ambiguity, missing context, conflicting sources and tool failures. Without written autonomy rules and evals, nobody can say how the agent will behave or when a person should step in.
It guesses
Behavior is judged by demos and gut feel, not by evals against real scenarios.
It hides what it doesn't know
When the agent is unsure or fails, users are not told, and a person cannot easily take over.
What I do
Calm, clear agent behavior.
I help teams decide what their AI agent should do on its own, when it should ask, and when a person steps in. Then I prove it with evals and a working prototype.
- AI agentsBehavior
- AI assistantsBehavior
- Agent evalsEvidence
- AI UX audits$1,500
- Behavior specsBlueprint
- Coded prototypesBlueprint
- Autonomy rules and approvalsBlueprint
- Trust and handoff designBlueprint
What you get
One senior product designer with a frontend background: behavior design, evals and code in one place.
Design, evals and code
Senior product designer and former frontend developer. You work directly with me. [PLACEHOLDER: years of experience]
Real prototypes
A working coded prototype of your agent's trust moments: evidence, approvals, action history and human takeover.
Async by default
Written updates and Loom walkthroughs instead of standing meetings. Fewer calls, clearer decisions.
Fixed scope, fixed price
Every engagement has a clear scope and price up front. The retainer can be paused or ended monthly.
Who this is for
Seed and Series A AI startups and AI dev shops with an agent or assistant that acts for users. You have engineers building it. You need to know it behaves well before more people rely on it.
Capacity
I take on up to five Blueprint clients a year, so each one gets my full attention.
What you provide
One product owner, staging access to your agent (or recorded transcripts), and a list of its tools and permissions.
The Agent Behavior Blueprint
Six weeks, grown in the open.
One agent workflow, $18,000 fixed. Each stage leaves you with a real deliverable, and a written update every week.
- Week 1Kickoff and mapDecisions, tools and permissions.
- Week 2Autonomy contractWhat it may do alone, what needs approval, what it must never do.
- Week 3Behavior specWording for hard moments: ambiguity, tool failure, takeover.
- Week 412-scenario evalRun against your real agent, with a pass/fail report.
- Week 5Coded trust-surface prototypeEvidence view, approvals, action history and takeover.
- Week 6ReadoutLoom and one 60 minute call. Go or no-go.
- Week 1Kickoff and mapDecisions, tools and permissions.
- Week 2Autonomy contractWhat it may do alone, what needs approval, what it must never do.
- Week 3Behavior specWording for hard moments: ambiguity, tool failure, takeover.
- Week 412-scenario evalRun against your real agent, with a pass/fail report.
- Week 5Coded trust-surface prototypeEvidence view, approvals, action history and takeover.
- Week 6ReadoutLoom and one 60 minute call. Go or no-go.
- Week 1Kickoff and mapDecisions, tools and permissions.
- Week 2Autonomy contractWhat it may do alone, what needs approval, what it must never do.
- Week 3Behavior specWording for hard moments: ambiguity, tool failure, takeover.
- Week 412-scenario evalRun against your real agent, with a pass/fail report.
- Week 5Coded trust-surface prototypeEvidence view, approvals, action history and takeover.
- Week 6ReadoutLoom and one 60 minute call. Go or no-go.
Step 1: Audit
$1,500. A review of your agent's experience and a sample of its outputs, with a prioritized fix list. Credited if you continue.
Step 2: Blueprint
Six weeks, $18k fixed. Autonomy contract, behavior spec, 12-scenario eval and a coded prototype.
Step 3: Decide
A pass/fail eval report, release-gate recommendations and a clear go or no-go.
Step 4: Keep it on track
Optional retainer: monthly eval reruns and fixes when behavior drifts after model or prompt changes.
Principles
What beautiful means here.
Soft on the surface, strict underneath. These rules shape every engagement.
Honest about uncertainty
Every answer separates what is known from what is guessed, with sources and dates. No decorative confidence scores.
Predictable under pressure
The agent can suggest and draft. Reliable systems calculate, approve and write, even when inputs get messy.
Respectful of people's time
Short, clear answers. High-stakes actions use clear approvals and buttons, not a typed "yes".
Knows when to hand off
When something fails or the stakes rise, the agent says so and hands control to a person. "The update was not published" beats a quiet error.
Pricing
Fixed prices: you know the cost before we start.
Start with an audit. Go deeper only when the need and the scope are clear.
AI UX audit
$1,500
Fixed price review
Find the trust, handoff and output problems in your AI agent. Credited in full toward the Blueprint. Delivered in about two weeks. [CHECK: turnaround]
- Review of your agent's experience
- Output quality review: sample outputs, sort the failures, trace each to a fix
- Prioritized fix list
- Written report and Loom walkthrough
- $1,500 credited toward the Blueprint
Flagship
Agent Behavior Blueprint
$18,000 fixed
Six weeks, one agent workflow
For AI startups and dev shops whose agent acts for users. Fixed scope, fixed price.
- Autonomy contract
- Behavior spec and sample wording
- 12-scenario eval against your real agent, with a pass/fail report
- Coded prototype of the trust moments
- Weekly written updates
- Kickoff and final review calls
- Go or no-go and release-gate recommendations
Async retainer
$4,500 to $6,000 a month
Keep it working
For Blueprint clients who want their agent's behavior checked every month as models and prompts change. Pause or cancel monthly.
- Monthly eval reruns
- Behavior drift fixes
- New behavior and prototype work, one request at a time
- Weekly written updates
- Async via Slack or email
2026 prices. One agent workflow per Blueprint.
Demo
See how I work.
Client case studies will be added as projects finish. Until then, here is the method applied to public examples.
[COMING SOON]
Teardown: how a public AI agent hands control back to a person
[PLACEHOLDER: demo to build, write after it exists] Draft: I ran [N] scenarios against a public agent, found where its handoff breaks, wrote the fix, and built a coded prototype of the better flow. Watch the Loom.
Worked example
A document assistant
Two source documents disagree about a deadline. The assistant shows both sources and their dates, flags the conflict, and asks the owner to choose. It can draft an update but cannot publish without approval. If publishing fails, it says "The update was not published."
FAQs
Questions and answers.
Tap a question to open it. Common questions about how I work.
Me, Zvonimir Perkovic. A senior product designer with a frontend background, based in Cleveland, Ohio.
If there's a fit, I send a one-page scope and a start date. Most teams start with the $1,500 audit.
You might later. A fixed Blueprint tests the agent's behavior now, before you commit to a hire or months of engineering.
Each engagement has a fixed scope and price. I take on up to five Blueprint clients a year, so each one gets focus.
An autonomy contract, a behavior spec, a 12-scenario eval report and a coded prototype, plus a go or no-go call.
Figma for design, and React with Tailwind for coded prototypes.
The Blueprint runs six weeks. On retainer, I work one request at a time with a written update every week.
Revision rounds are agreed in the scope. Weekly updates surface problems early.
No. It's a working prototype for testing and decisions. Building the production agent stays with your engineers.
[PLACEHOLDER: refund terms]
Yes. Evals run best against a staging version. With only recorded conversations, untestable cases are marked as not tested.
Building the production agent, training models, and safety certification. One agent workflow per Blueprint.
AI startups and dev shops with an agent or assistant that acts for users, and a team already building it.
Prove your AI agent behaves before you ship it
Let's make your agent something people trust.
Book a 20 minute call. Tell me what your agent does, and I'll suggest where to start. [PLACEHOLDER: booking link]
Next start: [MONTH]