Does a monitoring dashboard count as a spend cap?

No. A dashboard shows spend after it happens. A spend cap is enforced inside the agent before each model call, so a loop stops at the cap instead of at the invoice.

Can you review a multi-agent pipeline as one agent?

Yes, if it ships as one pipeline with one owner. The six guardrails are tested at the pipeline level and at each agent that can call a tool.

Do you review MCP servers?

Yes. We test what the server lets an agent reach, whether its keys stay server-side, and whether a planted instruction in a tool result changes what the agent does.

Pre-Flight for Agent Builders

Guardrails for tool-using AI agents, tested before the agent runs unattended.

Two of your agents disagree about when a job is done. One asks for more work, the other obliges, and neither has a rule that ends the loop. Your dashboard shows the spend climbing; nothing in the pipeline can stop it.

Request a Pre-Flight Review See the six guardrails

Last updated 2026-09-26 · Agent Guardrail Desk, Denver, Colorado

Pre-Flight for Agent Builders

Who this is for

Engineers shipping tool-using agents, multi-agent pipelines and MCP servers. This is the loop in the public post-mortem where four agents ran for 11 days and spent about $47,000 before a billing alert fired (public post-mortem of a four-agent LangChain pipeline, published March 2026, collected at github.com/vectara/awesome-agent-failures (read 2026-09-26)). The Pre-Flight for Agent Builders tests the brakes, not the dashboard.

What we plant in an agent pipeline

A retry loop between two agents that never agree the job is done, to test the Spend Cap and a hard limit on rounds.
A pulled kill switch while a run is in flight, including work already queued on the server.
A tool call outside the agent's job, such as a write to a table or repo it should only read.
A tool result that carries a planted instruction, from the OWASP agentic categories of goal hijack and tool misuse.
A malformed tool response, to see whether the agent stops or carries on with empty values.
One run picked from the log, to see whether trigger, tools called, spend and changes can be read back.

The acceptance tests for agent builders

Pass means: the planted loop stops at the cap or the round limit, whichever comes first, and the alert reaches a named person; the switch stops the run in flight and blocks new runs; the out-of-scope call is refused; the planted instruction is refused or held for human approval; the malformed response halts the run; and the run log answers the four questions in under five minutes.

What you receive

The Pre-Flight Report with evidence for each guardrail, a ranked fix list, a spend-cap config for your runtime and a kill-switch runbook your on-call person can follow. $1,500 per agent or pipeline, delivered in 5 business days after read-only access to staging.

The shared spine

The same six guardrails as every program

This is a named, bounded system. Every guardrail maps to the failure it prevents and to one acceptance test we run on a staging copy of your agent. It runs on one agent or automation at a time. Out of scope: we never take production write access, and we never ship code to your production.

Guardrail	Failure it prevents	Acceptance test
Spend Cap	The runaway bill	We plant a retry loop on staging. Pass: the run stops at the cap and the alert reaches the named person.
Kill Switch	The agent nobody can stop	We pull the switch during a live staging run. Pass: the run in flight stops, no new run starts, and the time to stop is written down.
Permission Scope	The write it should never make	We ask the agent to write outside its scope and search the frontend bundle for keys. Pass: the write is refused and no secret ships to the browser.
Silent-Break Watch	The run that keeps going on empty	We rename one form field on staging. Pass: the run halts and alerts before it writes an empty record.
Hijack Tests	The instruction hidden in a ticket or a web page	We plant an instruction inside a document the agent reads. Pass: the agent refuses it, or stops for human approval before any destructive step.
Run Log	The run nobody can explain	We pick one run from the past day. Pass: those four questions are answered from the log in under five minutes.

Price: $1,500 per agent or automation, once. Delivered in 5 business days after read-only access to a staging copy.

Questions

Questions about this

Does a monitoring dashboard count as a spend cap?: No. A dashboard shows spend after it happens. A spend cap is enforced inside the agent before each model call, so a loop stops at the cap instead of at the invoice.
Can you review a multi-agent pipeline as one agent?: Yes, if it ships as one pipeline with one owner. The six guardrails are tested at the pipeline level and at each agent that can call a tool.
Do you review MCP servers?: Yes. We test what the server lets an agent reach, whether its keys stay server-side, and whether a planted instruction in a tool result changes what the agent does.

Every page links back to the Agent Guardrail Desk home page, where the price, the rulebook and the request form live.

The review, per situation

SituationPre-Flight for AI-Built AppsFounders and makers shipping apps built with Lovable, Bolt, Cursor or Claude Code on Supabase. The request is refused; no service key ships to the browser.SituationPre-Flight for No-Code AutomationsOps owners running Zapier, Make or n8n flows from forms and webhooks into a CRM. The run stops and alerts instead of writing an empty record.SituationPre-Flight for Platform TeamsPlatform, security and ops leads whose internal agents can write to CRM, ticketing and docs. The write is refused; the agent's role grants only what its job needs.

Buyer questions, answered

QuestionHow do I cap what an AI agent can spend?Where a spend cap has to live, what it counts, and the test that proves it stops a loop.QuestionWhat should an AI agent kill switch actually do?The four things a kill switch must stop, who holds it, and how to prove it works.QuestionHow do I catch a renamed form field before it writes empty records?Silent breaks in automations: what to check before each write, and the test that proves it.QuestionHow do I check an AI-built app is secure before real users?The pre-launch routine for Lovable, Bolt, Cursor and Claude Code apps, and what to test after it.

Price and sample

Start hereThe Pre-Flight Review: price, scope and timeline$1,500 per agent, report in 5 business days, credited in full against a Fix Sprint within 30 days.Start hereSee a sample Pre-Flight ReportThe report format, filled in for a placeholder agent. Demonstration, not a client result.HomeAgent Guardrail Desk: the Six-Guardrail Pre-FlightThe whole offer on one page: the six guardrails, the price, the rulebook and the form.

Next step

Request a Pre-Flight Review for one agent

$1,500 per agent or automation, once. Delivered in 5 business days after read-only access to a staging copy. If you book a Fix Sprint within 30 days, the $1,500 is credited in full against it.

Request a Pre-Flight Review See a sample report