Your fantasy, executed.
We build AI agents for the work you can’t skip but never set out to do. They run it end to end; your people keep the calls that matter.
I wish Friday’s books closed themselves.
- Pull this week’s transactionsbank.export
- Match receiptsdrive.search
- Flag what doesn’t add uprules.check
- Review the 3 flagged itemsYour callA $1,240 vendor charge has no receipt.
- Send the close reportgmail.send
Sketch mode: scripted for this page. The real thing runs on your systems, with your approvals.
Agents we build
Necessary, repetitive, and not your core business — that’s where an agent pays off first.
-
Operations
Month-end closes, reporting, and the data that moves between the tools you already pay for. The agent does the routine; you sign off on the exceptions.
A first agent: the Friday close, from bank export to report.
-
Engineering workflows
Coding agents like Claude Code, set up inside your repository with your conventions, guardrails and review gates — so what they write is code your team merges.
A first agent: ready issues turned into reviewed pull requests overnight.
-
Customer-facing agents
Support and conversation agents grounded in your own docs, with a clean handoff to a person whenever money, risk or a frustrated customer is involved.
A first agent: overnight replies in every language your customers write in.
How we work
-
Sketch
One working session to map the fantasy: what you would hand off, what must never go wrong, and what “done” looks like.
-
Harness
We build the agent and everything around it — tools, permissions, tests, and the checkpoints where a person decides.
-
Prove
Every task leaves evidence you can check: a log, a diff, a receipt. Nothing is marked done on the agent’s word alone.
-
Hand over
Runbooks, monitoring and hands-on training, so your team can run and extend the agent. You own the harness; we stay on call if you want us.
Agents do the work.People make the calls.
Questions
What does Agentasy build?
AI agents for the work a company can’t skip but never set out to do: operations such as month-end closes, reporting and approvals; engineering workflows built on coding agents like Claude Code; and customer-facing support agents grounded in your own docs.
Do the agents make decisions on their own?
No. Agents run the routine end to end and stop at checkpoints where a person decides — approving an exception, issuing a refund, merging a change. Nothing irreversible happens on an agent’s word alone.
How does a project work?
In four steps. Sketch: one working session to map the work, what must never go wrong and what done looks like. Harness: we build the agent with its tools, permissions, tests and checkpoints. Prove: every task leaves evidence you can check, such as a log, a diff or a receipt. Hand over: runbooks, monitoring and training so your team runs it.
How do you test that an agent does the job right?
Every workflow has its own definition of “right”, so there’s no single score we promise. Before building, we work out with you which dimensions matter and how to grade them, using real past cases as the reference set. We backtest on your historical data before launch, rerun the full set whenever a model, prompt or tool changes, and keep tracking exceptions and review rates after go-live.
What can an agent access, and how are credentials handled?
Read-only and least-privilege by default, through a dedicated role or service account that you create and can revoke at any time. Anything that writes or reaches outside — a payment, an email, a merge — goes through a checkpoint a person approves. Credentials live in your platform’s secret manager, never in code or chat; sensitive fields are masked by role, and every tool call is logged.
Is our data used to train models, and where does it live?
We build on Anthropic (Claude) and Google (Gemini) under their commercial API terms, which exclude customer data from model training. We work as forward-deployed engineers inside the environment you provide, so your data stays in your systems and cloud accounts — we keep no copies. Only the fields a task needs are sent to a model, logs follow your retention and security policies, and our access is revoked when the engagement ends.
Who owns the code, prompts and accounts?
You do. Code, prompts, workflows, configuration and evaluation sets live in your repository from day one, and cloud and SaaS accounts are opened in your name and billed to you. We keep our methodology and reusable assets that contain none of your data — skills, scripts and templates — which is why each project starts faster than the last.
Can we switch to a newer model later?
Yes, and it’s worth planning for. A solid harness absorbs most of what changes between models, but a new model still has to pass the evaluations we agreed on before it goes live. That verification is most of the work of a switch, and we budget for it rather than assume it’s free.
Do the agents run on our own systems?
Yes. Agents run inside the tools you already use, with your permissions and your approvals. The demo on this page is a scripted illustration of that flow.
What kind of team is Agentasy?
A team whose people average twenty years in information technology, across production systems — telecom, certified payment platforms, POS and vision AI — and the back-office functions a company must run to stay compliant, such as finance, customer service and information security. Agentasy is based in Taipei and works with teams worldwide.
How do we start?
Describe the work you would hand off in the contact form on this page. We reply within one working day with how we would build it.
What would you hand off tomorrow?
Write it down in a sentence or two. We reply within one working day with how we would build it.
Prefer email? Write to hello@agentasy.ai.