Does it slow my agents down?

No. Rules are evaluated inside your own service with no network call, so there's no added delay, even for voice agents.

How long does setup take?

A few minutes. Install the SDK and add two lines to your entry point.

Which frameworks do you support?

LangChain, LangGraph, CrewAI, AutoGen, OpenAI Agents SDK, Google ADK, LlamaIndex, Haystack, Semantic Kernel and the Vercel AI SDK, plus direct calls to OpenAI, Anthropic, Gemini, Bedrock, Mistral, Groq and OpenRouter.

What can ZespanPilot do?

It answers questions about your production data in plain English, like "why did errors jump at 3pm?", and can take actions such as proposing a fix or creating a rule. Every action is role-checked, previewed and logged, and nothing ships without your approval.

Do you have a free plan?

Yes. The free plan includes 50,000 traces a month, with no credit card required.

Yes, on the Scale plan, including bring-your-own-keys and a dedicated SLA.

AI agent reliability

Your AI agent didn’t crash.
It just got it wrong.

Name: Zespan
Brand: Zespan
Availability: InStock

Zespan catches the silent mistakes your agents make in production, shows you what went wrong and why, and stops each one happening twice, with guardrails tested on your own traffic and fixes you approve.

Start free →Book a 20-minute walkthrough

Two lines to install · Free plan, no credit card · Self-hosting available

Works with the stack you already run

OpenAI

Anthropic

Google Gemini

AWS Bedrock

LangChain

CrewAI

LlamaIndex

Vercel AI SDK

OpenTelemetry Mistral

Mistral

Groq

AutoGen

Nothing errors. Nothing alerts. A customer finds out first.

Wrong answers with no error

The request succeeds, the logs look clean, every dashboard stays green. Then a customer tells you the agent got it wrong.

The same mistake, again

You fixed the prompt last week. The failure is back in a slightly different form, and nothing stopped it.

No way to prove what happened

When someone asks what the agent was allowed to do, and who approved it, there’s no clear answer.

Your monitoring and your guardrails were not built to catch this. Here is why.

Others show you the failure. Zespan stops it happening again.

Monitoring tools show you what happened, after it happened. Guardrail tools block content that is known to be bad. Neither knows that your policy says refunds stop at 30 days.

Capability comparison: observability tools, guardrail tools and Zespan
	Observability tools	Guardrail tools	Zespan
Shows what the agent did	Yes	No	Yes
Blocks known harmful content	No	Yes	Yes
Catches wrong decisions that look normal	After the fact	No	Yes
Turns a failure into a tested rule	No	No	Yes
Tests a rule on past traffic first	No	No	Yes
Runs inside your service, no added delay	No	Varies	Yes

Observability tools: LangSmith, Langfuse, Arize, Helicone. Guardrail tools: Lakera, NVIDIA NeMo Guardrails, Guardrails AI. Categories describe what each kind of tool typically does, and individual products vary.

See the detailed comparison →vs LangSmith vs Langfuse vs Arize vs Helicone

From first mistake to permanent fix.

Zespan catches the wrong answer your monitoring never saw, shows you why it happened, stops it happening twice, and keeps a person in charge of each step.

The four steps end to end. One, Catch: a run log with one wrong answer flagged among clean 200 OK runs. Two, Group: forty Promised a refund failures collapse into one issue. Three, Turn it into a rule: Never promise a refund we don't offer, applied to all agents and tried on last month, which would have caught 40. Four, Enforce: the wrong reply is stopped before sending and a correct one goes out. — The whole loop: catch the wrong answer, group the repeats, write the rule, enforce it. Illustrative example.

1 / 5

See the failures your monitoring misses.

Zespan checks every agent run, not only the ones that error. Wrong answers are flagged, grouped by what actually went wrong, and given an AI-written root cause, so a hundred alerts become the handful of issues worth fixing.

Alerts go to Slack and PagerDuty with the trace attached, and every run keeps its cost, latency and quality scores.

What you get: You hear about a mistake from Zespan, not from a customer.

A Zespan Detect view listing three open issues grouped by what went wrong, each with an AI-written root cause. Forty conversations promised a refund the company does not offer, because the refund prompt never states the 30-day limit. Twelve quoted the wrong delivery date because the order tool returns UTC. Seven skipped the hand-off to a person because the sentiment tool timed out. An alert went to Slack with the trace and root cause attached. — Failures grouped by what went wrong, each with a root cause and an alert. Illustrative example.

2 / 5

Test a guardrail on last month’s traffic before it goes live.

Most teams write very few guardrails, because shipping one blind can block real customers. Zespan replays a proposed guardrail against conversations you have already served. You see what it would have caught, what it would have wrongly blocked, and the real conversation behind every result.

Guardrails live as code in your own repository and are reviewed in pull requests. They move from dry run, to warn, to block, and a guardrail can only start blocking after it passes its test. Run one against last week with zespan policy test --against last:7d.

What you get: You ship guardrails with evidence, not hope.

A Zespan Prevent view replaying three rules on past traffic. Refund over policy limit would have caught 47 and wrongly blocked 2, personal data in reply caught 118 and wrongly blocked 0, medical advice boundary caught 6 and wrongly blocked 31 and needs work. Below, the real conversation behind a correct catch and behind a wrong block. — A guardrail replayed on past traffic, with the real conversation behind each result. Illustrative example.

3 / 5

The guardrail runs inside your own service.

A guardrail compiles into a bundle your SDK evaluates in-process. The check happens inside your own service, with no call to our servers and no proxy in the path.

That is why it adds no network delay, even on real-time voice calls where a round trip is not acceptable.

What you get: Protection on every request, with nothing rebuilt on someone else’s runtime.

Two request paths compared. With a proxy, the agent calls a proxy over the network before the model, adding a hop to every request. With Zespan, the policy check runs inside your own service between the agent and the model, with no network call in the request path, and Zespan syncs the rules ahead of time. — The guardrail is checked inside your service: no proxy hop, no network call in the request path. Illustrative example.

4 / 5

Risky actions wait for a person, and every decision is on the record.

An action with real consequences pauses instead of going through. A named reviewer decides, and the decision is recorded against them, not against the agent that asked.

Each audit entry shows whether a person, an API key or the system acted, and exports as SOC 2 evidence. How governance works.

What you get: A clear answer when someone asks what the agent was allowed to do.

An approval card for refund.issue, $4,000 over the $2,000 ceiling, held for reviewer finance-oncall with Approve and Reject buttons, above an audit trail of four entries: policy.blocked, approval.requested, approval.decided with actor user, and audit.recorded for SOC 2. — A risky action pauses for a named reviewer, and the decision lands in the audit trail. Illustrative example.

5 / 5

Fixes are proposed and tested. You approve them.

When an incident closes, Zespan proposes a fix and replays it against every failure that incident captured, so you see the pass rate before you decide anything.

Every approved or rejected fix becomes a permanent regression test built from a real production failure, so the same mistake is caught before it ships again. Nothing ships automatically.

What you get: A failure that stays fixed.

An incident card for RefundAgent, incident 412, awaiting approval. The proposed fix changes prompt v9 to v10, replacing Refunds are handled by the agent with Refunds are allowed within 30 days of delivery, otherwise offer a replacement. It was replayed against 41 captured failures: 39 pass and 2 regressions are flagged. Buttons read Approve to staging and Reject. Nothing ships until a person approves. — A proposed fix, replayed against the failures that caused it, waiting for a person. Illustrative example.

Tested on real production traffic.

In a pilot running 10 agents in production, Zespan grouped 104 raw failures into 10 real issues, and turned two of them into permanent tests so they can’t come back.

raw failures104

real issues10

permanent tests2

Read the full case study →

Built for teams running agents in production.

The more your agent does on its own, the more a silent mistake costs. And with a person in the loop, a reviewer can only catch what they are shown.

Customer support and refunds
Wrong promises, wrong refunds, missed escalations.
Financial services and insurance
Decisions and advice that carry financial and regulatory risk.
Voice agents
Wrong prices or commitments on live calls, where speed matters.
System integrators
Consistent governance across every client you deploy agents for.

Everything around the loop, built for agents.

The loop runs on a full toolkit for tracing, monitoring, cost and quality, all in one place.

Two lines of code.
Running in minutes.

Python and TypeScript. No forking, no architecture changes. Traces appear within seconds of the first agent run.

terminal

$ npm install @zespan/sdk
# or for Python
$ pip install zespan

agent.ts

import { zespan } from "@zespan/sdk";
zespan.init({
  apiKey: process.env.ZESPAN_API_KEY
});

Work on production failures from your editor.

Connect Claude Desktop, Cursor or any MCP client. Your coding agent can find what’s failing most, see the root cause, and draft a fix, with every action checked against your role.

claude_desktop_config.json

{
  "mcpServers": {
    "zespan": {
      "url": "https://api.zespan.com/v1/mcp",
      "headers": {
        "Authorization": "Bearer lqtp_..."
      }
    }
  }
}

Works in Claude Desktop, Cursor, and any MCP client over HTTP. Authenticates with a personal API key, not a project key — so every tool runs as you, under your role.

A Cursor agent session. The prompt reads: what's failing most in production right now, and fix it. Five tool calls follow. Read: zespan_issues_list finds 3 open issues, the top being checkout.refund with 41 occurrences. Read: zespan_issue_remediate finds the refund prompt never states the $2,000 ceiling. Write: zespan_prompt_create makes refund-agent v10 with the ceiling stated. Write: zespan_prompt_set_label labels v10 production after a regression check. Write: zespan_retroactive_eval_create queues yesterday's traffic for scoring. — One session: two reads to find the cause, three writes to ship and check the fix. Every tool checks your role first. Illustrative example.

Five of 35. The rest reach traces, prompts, issues, metrics & cost, datasets, dataset runs, evaluators, retroactive evals, sessions, projects, plain english — each one checked against your role before it runs.

Secure by design. Gated by default.

Built so a security review is a short conversation.

Talk to us →

Self-hosted deployment
Run Zespan in your own environment.
Your data stays yours
Rules run inside your service, and personal data can be redacted before it reaches a model provider.
Bring your own keys
Provider credentials never leave your control.
Approvals and audit trail
Every change is role-checked, previewed and logged.
Retention you choose
From 14 to 365 days, per project.
OpenTelemetry native
Send data to Zespan, your existing tools, or both.

Writing

See all posts →

Sep 27, 2026 · 5 min read

AI Agent Guardrails for Production: Beyond Security Filters

Security guardrails block prompt injection and PII leaks. They don't catch an agent confidently approving the wrong refund or calling a tool with the wrong argument. Here's the difference, and why production agents need both.

Sep 27, 2026 · 6 min read

How to Test an AI Guardrail Before Deploying It

Ship a guardrail blind and you can block real customers as easily as you catch real problems. Here's how to replay a proposed rule against your own past traffic first, and read the two numbers that tell you whether it's ready.

Sep 27, 2026 · 5 min read

Guardrails for Voice Agents and Real-Time AI: Enforcing Without Added Latency

A proxy-based guardrail adds a network hop to every request. For a real-time voice agent, that hop is the difference between a natural response and an audible pause. Here's why in-process enforcement is the only architecture that works for live calls.

Common questions

It catches mistakes your AI agents make in production, including wrong answers that don't trigger any error, groups them into real issues, and turns each one into a rule that stops it happening again.

They show you what your agents did. Zespan also turns a failure into a rule, tests it on your past traffic, and enforces it inside your service.

Your agents are making decisions right now. Do you know which ones are wrong?

Start free in minutes, or book a walkthrough on your own agents.

Start free →Book a 20-minute walkthrough

Free plan · No credit card · Two lines to install

Before and after. Without a fix in place, complaints reach 47: a refund promised in chat, an escalation to the customer's bank, a disputed charge. With Zespan, complaints stay at 0: a customer asks for a refund on a damaged order, the wrong reply is stopped before sending, and the agent offers a free replacement instead.

Your AI agent didn’t crash.It just got it wrong.

Nothing errors. Nothing alerts. A customer finds out first.

Others show you the failure. Zespan stops it happening again.

From first mistake to permanent fix.

See the failures your monitoring misses.

Test a guardrail on last month’s traffic before it goes live.

The guardrail runs inside your own service.

Risky actions wait for a person, and every decision is on the record.

Fixes are proposed and tested. You approve them.

Tested on real production traffic.

Built for teams running agents in production.

Everything around the loop, built for agents.

Two lines of code.Running in minutes.

Work on production failures from your editor.

Secure by design. Gated by default.

Writing

Common questions

Your agents are making decisions right now. Do you know which ones are wrong?

Others show you the failure. Zespan stops it happening again.

From first mistake to permanent fix.

See the failures your monitoring misses.

Test a guardrail on last month’s traffic before it goes live.

The guardrail runs inside your own service.

Risky actions wait for a person, and every decision is on the record.

Fixes are proposed and tested. You approve them.

Tested on real production traffic.

Built for teams running agents in production.

Everything around the loop, built for agents.

Two lines of code.Running in minutes.

Work on production failures from your editor.

Secure by design. Gated by default.

Writing

Common questions

Your agents are making decisions right now. Do you know which ones are wrong?

Your AI agent didn’t crash.
It just got it wrong.

Two lines of code.
Running in minutes.

Two lines of code.
Running in minutes.