AI agent development
that survives real users.

Copilots, retrieval assistants, tool-using and role-play agents on Claude and GPT, added to the PHP and Python web apps you already run or built fresh - built with the evals, guardrails and cost caps that keep them working past the demo. Prepaid in Hour Blocks of 6, 12 or 24 hours, and there's a live agent of ours you can talk to right now.

Copilots RAG & retrieval Tool use Evals & guardrails Cost control
MagpieFeed, a live product we built and run

Why work with AgiCAD

Within two business days

We read your message and reply - not a sales queue, not an account manager.

Hour Blocks from $180

Prepaid in Hour Blocks of 6, 12 or 24 hours. No fixed scope, no expiry on unused hours.

A live agent, shipped

Not a slide deck - a real LLM agent in production you can open and use in the next minute.

You own the code

Full documented source in your repository. No lock-in, no recurring licence to us.

Why demos work
and products don't.

Wiring a prompt to a chat box takes an afternoon. Everything below is why the second month is harder than the first.

It invents answers

A model with no grounding will confidently make things up. Retrieval over your own content, with citations back to the source and an explicit "I don't know" path, is what turns a demo into something support can stand behind.

Nobody can tell if a change helped

Prompts get tweaked on vibes and quietly regress. A held-out set of real cases with automatic scoring turns "it feels better" into a number you can compare between versions.

The bill is unpredictable

Unbounded context, retries and one viral day. Token budgets per request, per-user quotas, a global cap, caching and a smaller model for the easy path keep cost roughly proportional to value.

Tool use is unreliable

An agent that can call your systems will eventually call them wrongly. Narrow tool schemas, strict argument validation, idempotent operations and a confirmation step for anything destructive.

Untrusted text steers the agent

Anything the agent reads - a document, a web page, a support ticket - can contain instructions. Retrieved content is treated as data, privileges are kept minimal, and high-impact actions stay behind human approval.

It feels slow

Ten seconds of silence reads as broken. Streaming, optimistic UI, parallel retrieval and caching of repeated work change the perceived speed more than any model swap.

Agents, not just API calls.

The interesting part is never the prompt. It's retrieval, tools, evaluation, guardrails and a cost model that doesn't surprise you.

Product copilots & assistants

An assistant inside your product or workflow that knows your domain - answering, drafting and acting on your data instead of guessing.

  • In-app chat copilots
  • Drafting & summarisation
  • Support and onboarding assistants
  • Streaming responses with citations
  • Conversation memory & context management
  • Claude, GPT or open models

Retrieval (RAG) over your content

Grounding the model in your documents, tickets and database so answers come from your material and can be traced back to it.

  • Ingestion & chunking strategy
  • Embeddings and vector search
  • Hybrid keyword + semantic retrieval
  • Re-ranking & citation of sources
  • Incremental re-indexing
  • Permission-aware retrieval

Tool-using & workflow agents

Agents that do things - call your API, file the ticket, update the record - with the constraints that make that safe.

  • Narrow, validated tool schemas
  • Idempotent and reversible operations
  • Human-in-the-loop approval steps
  • Multi-step planning with step limits
  • Retries, timeouts & fallbacks
  • Full audit trail of actions taken

Evaluation, guardrails & observability

The part that decides whether you can change anything later without fear.

  • Eval set built from real cases
  • Automatic scoring & regression checks
  • Input and output guardrails
  • Prompt-injection hardening
  • Prompt, latency & cost logging
  • Dashboards and alerting

Cost control & scaling

Making the unit economics work before the traffic arrives, not after the invoice does.

  • Token budgets per request
  • Per-user quotas & global caps
  • Prompt and response caching
  • Model routing (small model first)
  • Batching & async processing
  • Cost-per-conversation reporting

From idea to an agent
you'd let customers use.

The fastest way to find out whether this works for you is a small, honest prototype - not a six-week discovery phase.

1

Tell us the job

What the agent should do, who uses it, and what data it can draw on. We will also tell you when a plain search box or a rules engine would beat an LLM.

2

Pick an Hour Block

Prepay a block of hours - 6, 12, or 24. The block is the budget; we work until it's used, or you top up. Hours don't expire.

3

Prototype against real cases

A working agent on your actual data, measured against an eval set built from real examples - so "is it good enough?" gets an answer instead of an opinion.

4

Harden & hand over

Guardrails, cost caps, logging and monitoring in place, source in your repository, and a written note on how to change the prompts without breaking the evals.

Talk to an agent
we built.

The fastest way to judge agent work is to use one. Parley is a real AI feature in production inside SignedDocsRepublic - no login, no signup.

Live demo

Parley - AI Communication Coach

A conversational agent that runs live LLM role-play: you practise a difficult conversation, it plays the other side and gives feedback afterwards. It ships with the unglamorous production parts - a per-browser usage quota, a global cost cap so one viral day can’t blow the budget, streaming responses, no-login access and streak tracking. Same engineering we would bring to your agent.

Open the live demo Read the case study

Internal tool · FastAPI

SEO Wander

An internal tool built around LLM output - structured generation, validation of what comes back, and a cost model that had to survive our own bill.

Read the case study
learncoachassist.com
LearnCoachAssist spaced repetition study interface

LearnCoachAssist

Laravel · AI

AI-assisted content generation inside a subscription product, with automated moderation, validation and tests on what the model produces.

Read the case study
Vision · audit

Vision-LLM image audit

A write-up of auditing 1,708 images with a vision model: prompt design, disagreement handling, and what the failure cases actually looked like.

Read the write-up

Everything above is our own work, designed, built, deployed and run in-house. More AI in production: the assistants inside NajNajNovije and UslugeHR. Client work stays under NDA, so what we show is what we run ourselves.

Prepaid in Hour Blocks.

Three block sizes. An Hour Block is a bucket of time, not a fixed scope - pick the block that fits the size of the job.

Starter

Starter - 6 hrs

$180

Typically enough for a focused prototype, an eval pass, or a cost-control fix.

Standard

Standard - 12 hrs

$360

Typically enough for a working agent with retrieval over your content.

Pro

Pro - 24 hrs

$720

Typically enough for a production agent with tools, evals, guardrails and monitoring.

Hour Blocks are sold at $30/hr with no discount. Unused hours don't expire - apply them to the same project, or roll them into a top-up Hour Block.

Frequently asked.

What's an Hour Block - and what if our job takes more or fewer hours?

An Hour Block is a prepaid block of hours: Starter (6h, $180), Standard (12h, $360) or Pro (24h, $720). It is a bucket of time, not a fixed-scope deliverable. If the work finishes before the hours are used, the leftover rolls forward to the same project or into a top-up Hour Block - hours do not expire. If the work runs over, you buy another Hour Block.

Which models do you build on?

Mostly Claude and GPT through their APIs, and open-weight models where privacy, cost or latency argue for self-hosting. The application is written so the model is swappable: prompts, retrieval and tools live in your code, not inside one vendor's agent builder.

What does the API usage cost, on top of your time?

That is billed by the model provider directly to your account, not through us - so you see the real number. Part of the build is making it predictable: token budgets per request, per-user quotas, a global cap, caching, and routing easy requests to a cheaper model. We will give you a cost-per-conversation estimate from the prototype's measured usage.

How do you stop it making things up?

Grounding and measurement. Answers are retrieved from your own content and cited back to the source, the agent is given an explicit way to say it does not know, and outputs are validated against a schema where they feed another system. Then an eval set built from real cases catches regressions when a prompt changes.

Is it safe to let an agent touch our systems?

It can be, with constraints: narrow tool schemas with validated arguments, idempotent and reversible operations, a hard limit on steps, human approval for anything destructive, and a full audit trail. Retrieved text is treated as data rather than instructions, which is the main defence against prompt injection.

What about our data - privacy and training?

Your data stays in your infrastructure and your provider account, and the major providers do not train on API traffic by default. Where that is still not acceptable, the same architecture runs against a self-hosted open model. We will sign your NDA or use a simple mutual one, and no client is ever named on this site.

Do we own the code?

Completely - the application, prompts, eval set and infrastructure config, in your repository, documented. Nothing is locked to a platform only we can operate.

How long until we can try something?

A prototype on your real data usually lands inside the first Hour Block. It is deliberately the cheapest way to find out whether the idea works, and we would rather tell you it doesn't after six hours than after six weeks.

related

Describe the agent.

A few sentences is plenty - what it should do, who uses it, and what data it can draw on. We read every message and reply within two business days with an honest take and which Hour Block fits.

Prefer email?

[email protected]
Please enter your name
Please enter a valid email
Tell us a little about the agent
Thanks - we've got it.

Expect a reply from us within two business days.

Hour Blocks from $180reply within two business days Start a project