ForthClover
Agent Build · Custom and customer-facing agents

Prove the agent first. Then build it for production.

Some agents can't be a two-week experiment. They talk to your customers, touch money or work across several systems. We build them in two stages: a fixed-price proof on your real work with a go/no-go at the end, then — only if the numbers hold — a production build in your cloud with evals, guardrails and monitoring.

Compare the two stagesWhat moves the price
Proof stage
$9,500
fixed · 3–4 weeks · one agent
Production stage
$24k–$40k
fixed · 6–8 weeks
The gate
Go / no-go
Stop after the proof and owe nothing more
Runs in
Your cloud
Infrastructure as code, in your repo
Two stages, one gate

Pay to find out if it works. Then pay to make it last.

The proof stage answers one question with evidence. Production only starts if the answer is yes, and it's quoted on what the proof measured.

Stage 1

Proof stage

$9,500

fixed · 3–4 weeks · one agent, up to 3 integrations

It answers: Does this agent do the job on our real work, measured against our own baseline?

What's included

  • Baseline measured on your real work in week one
  • Success metric written and signed before we build
  • One agent, connected to up to 3 of your systems
  • An eval set built from your real examples, hard cases included
  • A live trial with your team reviewing every output
  • Go/no-go readout with the measured result

You leave with

A measured go/no-go decision, and a fixed Production quote if it's a go.

Stage 2 · only after a go

Production stage

$24,000–$40,000

fixed · 6–8 weeks · priced by integrations and users

It answers: Can it run safely for real users, every day, without someone watching it?

What's included

  • Evals as CI gates on every change to prompts, tools or models
  • Guardrails: tool permissions, approval steps, input and output checks
  • Monitoring and OpenTelemetry tracing, with alerts on quality, cost and errors
  • Infrastructure as code (Terraform or CDK) in your cloud account
  • Security review against the OWASP Top 10 for LLM Applications (2025)
  • Runbook and handover sessions with your team

You leave with

An agent in production in your cloud, and a team that knows how to run it.

The gate between stages

At the end of the proof stage you get a written readout and make the call. A no-go ends the engagement there.

  • The result against the success metric you signed
  • Eval scores on your hardest cases
  • Running cost per task at your expected volume
  • Risks, and what production has to fix
  • A fixed Production quote within $24,000–$40,000
  • Our honest recommendation, including “stop”
Price guide

What moves a Production quote between $24,000 and $40,000.

Production is priced by how much the agent connects to and who depends on it. No hourly billing, no multipliers: the quote is fixed at the readout, once the proof has shown what the work involves.
$24,000$40,000

Beyond

What moves a production quote
What we price onToward $24,000Toward $40,000Custom quote
IntegrationsSystems the agent reads from or writes toThe proof stage's systems, with good APIs or existing MCP serversMore systems, or ones that need a custom connector or MCP serverMany systems, or legacy ones with no usable API
UsersWho relies on it day to dayOne internal teamSeveral teams, or your customersSeveral products or customer groups, each with its own rules
ChannelsWhere people reach itOne channel, such as Slack, Teams or a web chat widgetSeveral channels, such as web chat, email and Slack togetherVoice and phone, or a channel that needs its own app
ActionsWhat it is allowed to doReads and drafts; people approve and sendWrites to your systems, with approval rules per actionSeveral agents handing work to each other
ComplianceRules your data has to followStandard business data and your usual access controlsPersonal data, single sign-on, audit trails, data-residency limitsRegulated data (for example HIPAA or PCI) or formal certification work
Bigger than this? Several agents, a platform for many teams, or regulated data becomes a custom quote, scoped on a call. Model and cloud usage is billed at cost to your own accounts and isn't part of the build price.
What we build

The agents teams bring to an Agent Build.

Each one starts with the same proof stage. The systems, the checkpoints and what we measure change with the job.

Customer-facing support agent

Answers customers in chat or email from your own docs, takes simple actions, hands off the rest.

Today

  • The same questions answered by hand, all day
  • Customers wait overnight for an order status
  • Refunds and changes need a person for every step

With the agent

  • Common questions answered from your help center, with sources
  • Order and account lookups done through scoped, read-only tools
  • Refunds and changes above your limit wait for a person's approval

Example run

Illustrative
  1. 14:02Chat: “My order still hasn't arrived after 9 days”
  2. 14:02Order #10482 looked up through the order tool (read-only scope)
  3. 14:02Carrier shows a depot exception — status and options sent, sources attached
  4. 14:03Customer asks for an $86 refund — above the auto limit, sent for approval
  5. 14:05Approved by an agent; refund issued, ticket summarised and closed

Systems it touches

Zendesk or IntercomHelp centerOrder system or ShopifyStripeCRM

We measure

  • Resolved without a hand-off
  • Wrong-answer rate on the eval set
  • Customer satisfaction on agent chats

Human checkpoint

Money moves only after a person approves. Upset or unclear customers go straight to your team.

Agents we've shipped

GrayHair Software · AI Agents

AI agent on AWS delivering instant access to postal-industry knowledge

  • 100+ employee inquiries handled by the agent in the first week of rollout
  • Replaced manual Confluence + USPS-manual lookups with sub-second answers in Slack
  • Daily-synced knowledge base keeps responses aligned with current postal regulations
Read the case study

AppEvolve · RAG

Slack-integrated AI chatbot on AWS for more accessible internal documentation

  • ~60% reduction in time engineers spent searching internal documentation
  • Nightly Notion sync keeps the knowledge base aligned with the latest standards
  • 100% adoption across the engineering team within the first month
Read the case study
How it's built

Your systems stay where they are. The agent gets only the access it needs.

Every tool the agent can call has its own scope. Risky actions stop for a person. Every run is traced and scored, so you can see what it did and why.

  1. 01

    Your systems

    Where the work and the data live

    CRMHelp deskERP / accountingWikis & drivesDatabasesSlack, Teams or web chat

  2. 02

    MCP servers & APIs

    Least privilege, per tool

    MCP servers with OAuthOfficial APIsWebhooks & queuesScoped credentials

  3. 03

    Agent runtime

    Permissions and approvals built in

    LangGraph or provider SDKPer-tool permissionsApproval stepsState & memoryRetries & fallbacksInput & output guardrails

  4. 04

    Models

    Routed by task and cost

    AnthropicOpenAIGoogleOpen-weightRouting & caching

Across every layer

Evals & tracing

Evals run as CI gates; OpenTelemetry traces every run and tool call.

Across every layer

Human approvals

Actions above your thresholds wait for a named person, and every decision is logged.

Week by week

Four weeks to a decision. Up to eight more to production.

Shown at the longest: a 4-week proof and an 8-week production stage. Production starts when you say go.

Proof stage · 3–4 weeks

  1. W1Baseline & success metricMeasured on your real work; metric signed
  2. W2–3Build the agentWorkflow, up to 3 integrations, eval set
  3. W3–4Live trialYour reviewers on real work; readout

Production stage · 6–8 weeks

  1. W5–6Infrastructure as codeEnvironments and CI/CD in your cloud
  2. W5–7Harden & integrateRemaining integrations, permissions, approvals
  3. W7–9Evals in CI & monitoringCI gates, tracing, dashboards, alerts
  4. W9–10Security reviewPrompt injection, permissions, data egress
  5. W10–11Staged rolloutFirst users, then everyone
  6. W12HandoverRunbook and sessions with your team
Who does what

What we need from you, and what we own.

You provide

  • An owner for the agent, about three hours a week
  • Access to the systems it uses, sandbox or test accounts first
  • 30–50 real examples of the work, including the hard cases
  • Reviewers for the live trial and a sign-off on the metric
  • A cloud account and an IT or security contact for production

We provide

  • Senior engineers on both stages, start to finish
  • The agent, its integrations, permissions and approval steps
  • Baseline, eval suite, tracing and monitoring
  • Infrastructure as code, the security review and the runbook
  • Handover sessions so your team can run it

Not included

  • Model and cloud usage (billed to your accounts)
  • Running and on-call after handover (see AI Operations)
  • More than one agent per build
  • Large data clean-up or migration projects
  • Formal compliance certification, such as a SOC 2 audit
Is this you?

Built for agents that have to hold up in front of real users.

A good fit if…

  • You know the job you want the agent to do, and you can write it down
  • It talks to customers, touches money or spans several systems
  • The systems have APIs, MCP servers or exports we can use
  • You want evidence before you commit to production spend
  • Someone on your side will own the agent after launch

Not the right start if…

  • You haven't picked a workflow yet — start with the AI Agent Pilot
  • You want a ready-made sales, documents or support agent — see Packaged Agents
  • You need an MCP server so outside agents can use your product — see Agent-Ready Integration
  • You need many agents at once — we'd still prove one first
Where it sits

Between a quick pilot and running it for the long haul.

Many builds follow a AI Agent Pilot on a simpler workflow. Before a customer-facing launch, an Agent Security Review adds a one-week adversarial test. After handover, AI Operations keeps it measured and current. How we build every system
  1. Step 1AI Agent Pilot$3,000fixed · 14 daysDetails
  2. Step 2 · you are hereAgent BuildFrom $9,500proof stage, then production from $24,000
  3. Step 3Agent Security Review$3,500pre-launch check · 1 weekDetails
  4. Step 4AI OperationsFrom $990/moafter handover · 3-month minimumDetails
FAQ

Questions we hear on fit calls

Why two stages instead of one price?
Because the first question decides everything: does the agent work on your real data? The proof stage answers it for $9,500 before you commit to production. If it works, production is quoted on what we measured, not on guesses.
What if the proof stage says no-go?
You stop there and owe nothing for production. You keep the readout, the baseline, the eval set and the code we wrote. Often the readout points to something smaller first, such as cleaning a data source or starting with a simpler workflow.
How is the production price set?
It's a fixed quote, given at the go/no-go readout, between $24,000 and $40,000. It moves with the number of integrations, the users and channels, what the agent may do and your compliance needs. Anything bigger becomes a custom quote.
Which models and frameworks do you use?
Whatever passes your evals at the lowest cost: models from Anthropic, OpenAI or Google, or an open-weight model, with routing so routine steps go to smaller models and caching where it helps. Orchestration is usually LangGraph or a provider's agent SDK. Tools connect through MCP servers or official APIs, and every run is traced with OpenTelemetry.
Where does it run, and who owns it?
In your cloud account (AWS, Azure or Google Cloud), deployed as infrastructure as code from your repository. You own the code, prompts, eval set and infrastructure. Model and cloud usage is billed to your own accounts.
Is a security review included?
Yes. The production stage includes a review of what we built: tool permissions, prompt-injection tests, data egress and secrets, checked against the OWASP Top 10 for LLM Applications (2025). For a separate one-week adversarial test before launch, or for agents someone else built, there's the Agent Security Review ($3,500).
What happens after handover?
Your team runs it with the runbook, or we keep running it under AI Operations, from $990/mo: monitoring, evals on every model update, cost tuning and a monthly report.
Proof $9,500 · Production $24k–$40k

Find out if the agent works before you pay to ship it.

A free 45-minute fit call to scope the agent, the systems it needs and the number that would make it worth building.

Book the fit call