All servicesService 01 / 03Agents / Tool use / Human review

Agentic systems

I build AI agents that can look up information, choose which tool to use next, and hand a decision to someone on your team.

Typical length
4 to 8 weeks
Format
Build engagement
See what is included
Spheres joined by branching paths on a dark panel, shown as a dot mesh.
(01) Overview

Overview


An agent is useful when the next step depends on what it finds. It reads a request, gathers the records it needs, picks a tool and proposes an action. The hard part is not the model. It is deciding what the agent may touch, where it must stop and how a person checks its work.

I design those boundaries with your team, then build the agent and the screen that sits beside it. Every action it proposes shows the sources it used, so a reviewer can approve, edit or take over in a few seconds.

(02) Example use

An assistant for your support desk


A customer asks why their bill changed. An agent can check the invoice, look up the relevant policy and draft a reply with the evidence attached. If the customer needs a refund, the agent sends the amount and its reasoning to a person for approval.

  1. 01Customer request
  2. 02Invoice lookup
  3. 03Policy check
  4. 04Draft reply
  5. 05Human approval
(03) Deliverables

What you receive.


D/01

A connected agent

An agent for one defined job, connected to the tools and records it needs. Together we decide what it can read, what it can change and when it must stop.

D/02

A review screen

Your team can inspect the source material, edit a proposed action and approve it or take over.

D/03

Test conversations

A set of realistic requests, including missing records and failed lookups, so you can check how the agent handles an unfamiliar situation.

Further reading: Anthropic on building effective agents
(04) Process

Four steps, one task.


  1. 01

    Map

    Trace ten real requests from arrival to resolution and mark each decision.

  2. 02

    Bound

    Agree on the tools, the permissions and the stop conditions.

  3. 03

    Build

    Ship one complete path with a review screen your team can use.

  4. 04

    Test

    Run ordinary and awkward cases, fix failures, agree on release.

Where to startBring a task where the next step depends on what someone discovers. A few example requests and the tools your team uses are enough to map the first version.

(05) Lab

Route a support request.


neu-lab / agentic-systemsRuns locally

Route a request.

This example uses keyword rules to show one handoff decision. Choose a sample or enter a request to see where it goes and why.

Destination

Run the request to see its destination and the rule that sent it there.

(06) Questions

Common questions.


Yes, for actions you agree are safe and easy to undo. Anything that changes money, accounts or personal data goes to a person first.

The agent reports what it tried and what is missing, then hands the case to a person with that context attached.

Anything with an API or a structured export. We start with the two or three tools the task actually needs.

Start a brief for agentic systemsBring one taskNext / 02Workflow automation