← all services

agentic engineering / production ai systems

agentic engineering.

autonomous agent systems shipped inside your product, state, memory, tool use and guardrails, not a chatbot bolted on top.

for bangkok fintech and saas ctos, and scale-up founders who need an ai agent to be a real product feature, not a demo. we build multi-step agents, retrieval pipelines and ai-native product surfaces as senior software engineering, shipped into a next.js and supabase codebase you own from day one, not a no-code workflow layer you rent forever.

book an architecture scoping (usd 4,750 / thb 171,000)

01 . what it is

the service in detail.

stack: next.js + supabase + claudepattern: state, memory, tool use, evalsownership: your repo, day onelanguages: en + th where relevant

a chatbot is a text box wired to an api key. an agent is a system: it holds state across steps, decides which tool to call and when, retries when a call fails, and stops itself when it is about to do something it should not. the gap between a claude prompt in a demo and an agent running inside a production next.js app is the same gap between a spreadsheet macro and an actual backend service. an autonomous support agent that reads a ticket, checks three internal systems and drafts a reply a human approves in ten seconds instead of ten minutes. a document-processing pipeline that ingests thai and english invoices, extracts line items with a retrieval layer over your own product catalog, and writes structured rows to postgres with a confidence score attached. a research agent that runs six searches, cross-checks sources and returns a memo instead of a wall of links. that is the shape of work that lives in this service.

prompt engineering gets a demo to work once. software engineering gets it to work the four-hundredth time, at 2am, when the upstream api times out and the agent has to retry with backoff instead of inventing an answer and shipping it to a customer. that is why this work sits with senior engineers, not prompt specialists: the hard part is the state machine that tracks where an agent is inside a multi-step task, the tool-use layer that gives claude a typed, permissioned surface instead of raw api keys, the eval suite that catches a regression before a client does, and the guardrails that stop an agent from taking an action it was never authorized to take. anthropic's tool-use apis handle the model-calling part well. the engineering around that call, memory design, retry logic, cost controls, evals, is what decides whether an agent ships or gets pulled after week two.

shipped into your codebase, not ours. same discipline as our custom apps service: the agent lives inside your next.js and supabase codebase, github repo under your org from the first commit, supabase project under your billing, claude usage on your own anthropic console. if you already run an internal tool on next.js and supabase, we add the agentic layer to that codebase directly, auto-triage on a support queue, auto-draft on a document type, auto-reconcile on a finance export, instead of rebuilding the platform around a bolt-on tool.

most 'ai automation' pitches in bangkok in 2026 are a zapier or make.com flow with a claude or gpt node stitched in: no state, no memory, no eval suite, no ownership, and no real answer for what happens when the flow hits an edge case at 3am. that is fine for a one-off internal automation. it is not a foundation for a product feature a paying customer touches. agentic engineering is the software-engineering version, the same architecture rigor as a payment flow or an auth system, because an agent that reads customer data, drafts customer-facing replies or triggers a refund is exactly that kind of surface, not a workflow toy.

the engagement starts with a paid technical discovery, not a free strategy call and not a refundable scoping gimmick like claudekick's. discovery produces a written architecture spec (state model, tool inventory, memory strategy, failure modes) and an eval plan (the concrete test cases and pass thresholds the build gets measured against) before a line of the agent loop gets written. that document is yours whether or not you build with us.

02 . who it is for

bangkok teams that get stuck here.

archetype 01

nok, a bangkok fintech cto shipping a support agent

series a lending platform, 45 employees, support queue growing faster than headcount. the board wants an ai support agent live as a shipped product feature, not a chatbot bolted onto the marketing site. nok needs the agent to read a ticket, pull account status from postgres, check three internal systems, and draft a reply a human reviewer approves inside ten seconds, with an eval suite that proves accuracy before it goes live and catches drift after.

archetype 02

kade, a saas founder adding an agentic layer

internal ops tool already running on next.js and supabase, three years old, thousands of rows moving through it by hand every week. kade does not want a rebuild, wants auto-triage on the intake queue and auto-draft on the weekly report, shipped into the codebase that already exists. the agent has to respect the rls policies and the audit log already in place, not sidestep them.

archetype 03

aew, ops director at a bangkok logistics or dtc seller

multi-warehouse fulfilment across three marketplaces (shopee, lazada, a shopify own-site), reconciliation done by two people in a spreadsheet every friday. aew needs a multi-step agent wired into the existing ops app: read the marketplace order feed, cross-check stock across warehouses, flag mismatches, draft the reconciliation entry, and stop and escalate to a human the moment a number looks wrong. not a claude chat window someone forgets to check.

03 . what you receive

concrete delivery, no abstractions.

04 . timeline

the studio method, four phases.

phase 0

discover and architect.

1 to 2 weeks

paid technical discovery. stakeholder interviews, existing codebase review, state model design, tool inventory, memory strategy, failure-mode mapping. output is the architecture spec and the eval plan, signed before build starts.

phase 1

build the core agent loop.

weeks 1 to 3 of build

the state machine, the model-calling layer, the first tool integrations stubbed. weekly demo. this is where the agent starts completing a multi-step task end to end, even before the guardrails are hardened.

phase 2

integrate tools and guardrails.

weeks 4 to 6 of build

full tool inventory wired (internal apis, database reads and writes, external services), permission boundaries enforced per tool, retry and backoff logic, the guardrails that stop an unauthorized action before it executes.

phase 3

eval, harden, and ship.

weeks 7 to 9, plus 90 days

eval suite run against the full test set, observability live, production cutover, handover documentation delivered. 90-day post-launch care begins, tuning against real traffic instead of synthetic test cases.

05 . pricing

transparent. no hidden lines.

price by scopefixed per phase

paid technical discovery: usd 4,750 (thb 171,000), fixed fee, credited to the build if you proceed. covers stakeholder interviews, codebase review, the architecture spec and the eval plan. no refund mechanic like claudekick's scoping, because the deliverable here is bespoke architecture work, not a productized package. build runs from usd 24,500 (thb 882,000) per phase, fixed price, same discipline as our custom apps builds: 6 to 14 weeks end to end depending on how many tools and integrations the agent touches. thb shown at approximately 36 to the dollar, bot reference rate at invoice date is what actually bills. usd or thb both accepted. terms: 50/50 for engagements under usd 25k, 30/40/30 for engagements usd 25k and above. stripe usd, wise usd, or thai bank transfer in thb.

06 . proof

cases that carry this service.

07 . questions

what bangkok buyers ask first.

01

how is this different from just using chatgpt or claude directly?

chatgpt or claude in a browser tab is a person typing a prompt and reading an answer. what we build is a system: the agent holds state across a multi-step task, calls your internal tools and apis through a permissioned layer, retries when a call fails, and gets measured against an eval suite before and after it ships. the model call is a small fraction of the actual engineering. the rest is the state machine, the guardrails, the observability and the retrieval layer, none of which exist when someone opens a chat window.

02

how do you prevent an agent from doing something wrong?

guardrails are designed during discovery, not bolted on after an incident. every tool the agent can call is permissioned and scoped: an agent that drafts a refund cannot also execute one without a human approval step, unless that specific action is explicitly authorized in the architecture spec. retries have backoff and a hard stop. every run is logged, so an anomaly is visible inside minutes, not discovered by a customer complaint. the eval suite runs the agent against known edge cases before every deploy, including the failure cases it is expected to handle correctly.

03

do we own the code?

yes, same as custom apps. github repo under your org from the first commit, supabase project under your billing, claude usage on your own anthropic console. the agent ships inside your codebase, not a separate tool you rent from us. if we disappeared tomorrow, your engineers could keep running it.

04

can this run alongside our existing ai-integration or claudekick setup?

yes, the two are complementary, not competing. claudekick automates operational workflows on top of the tools you already run: your accounting software, your line oa, your crm. agentic engineering ships agent capability inside your product itself, a feature your customers or your internal team touch inside the app you own. a fintech client could run claudekick for the finance team's monthly close and a support agent built under this service inside the actual product, both on the same claude account, no conflict.

05

what does 'eval suite' mean in practice?

a documented set of test cases the agent has to pass before it ships and before every future deploy: known inputs, expected behavior, pass or fail thresholds, edge cases and failure modes included. wired into ci so a prompt change, a model version upgrade or a new tool integration gets checked automatically instead of relying on someone noticing the agent got worse. this is the difference between an agent you can trust in production and a demo that worked once.

06

what stack do you use?

next.js and supabase as the default application layer, same as our custom apps builds. claude for the model layer, using anthropic's tool-use apis for the agent's actions. postgres with pgvector or a dedicated vector store for retrieval where a rag pipeline is part of the design. sentry and supabase logs for observability, ci-wired eval runs on every deploy. aws region of your choice or fly.io when latency or data residency matters.

08 . further

not quite the right fit?

ready to scope the architecture?

agentic engineering starts with a written spec.

the paid technical discovery produces the architecture spec and the eval plan before a line of the agent loop gets written. fixed fee, credited to the build if you proceed. line official account fallback for thai-first contact.