Skip to content

Labor · Agent systems

Specialized agents that hand off tasks with intent

Instead of wiring every service to every other service, agents describe their capabilities on a card and find matching counterparts at runtime. Plug in a service yourself, hand over a task and play through a system in motion.

Practical setup 01Test bench · A2A principle

Finding capabilities instead of hardwiring connections

Hardwired integrations need individual changes with every extension. In modular agent systems, new capabilities should be discoverable on demand, without fixing every connection in advance.

Test bench

Plug in a new service, then hand over a task.

Connection

New services

The service's card

No new service plugged in yet.

Hand over a task

Services in the network

6

plugged in and existing

Changes

0

places that had to be changed

Handed over

0

tasks with a counterpart

No counterpart

0

went to a person

Example with invented services. Cards with skills and finding counterparts through those cards follow the A2A protocol. The number of changes counts places, not effort.

How the setup works Pick the connection and plug in a new service. Via the directory it publishes a card, hard-wired every user needs its own line. Then hand over tasks and see who finds the counterpart.
Scope and limitations

Not part of this demo: The network shows roles and paths, no response times and no count of production agents. The changes count places, not effort.

Cards with skills and finding counterparts through them follow the A2A protocol. The services are examples, not a customer system.

Practical setup 02Test bench · A2A states

A handover has a state

Whoever hands a task to another agent needs to know whether it arrived, whether it is being worked on, whether a question is open or whether it failed. Without that, all that is left is waiting.

Test bench

Pick a course and hand over the task.

ValidationCheck invoice R-2041 from ItalyForeign invoices
    Course

    Messages

    0

    between both agents

    State changes

    0

    since the handover

    Your answers

    0

    to questions

    Example process with invented values. The seven states and the question via input-required follow the A2A protocol.

    How the setup works Pick a course and hand over the task. The current state lights up at the top, the exchange is shown below. When a question comes, you answer it yourself, and you can cancel at any time.
    Scope and limitations

    Not part of this demo: The exchange is scripted. A real specialist answers with its own texts and takes time to do so.

    The seven states and the question follow the A2A protocol, which our public agent also speaks.

    Interactive simulation

    Play through an agent topology yourself

    Start different process types and watch which specialized stations get involved. The simulation shows paths, sequences, and queues, but no measurements from a live system.

    Agent topology: 13 building blocks in 5 layers, 13 connections derived from the processing paths.
    L0 · 3 BUILDING BLOCKSL1 · 2 BUILDING BLOCKSL2 · 3 BUILDING BLOCKSL3 · 4 BUILDING BLOCKSL4 · 1 BUILDING BLOCKCPCustomer PortalISInternal SystemsPAPartner AccessCLClassificationVLValidationRTRouterKBKnowledge BaseDERulesetPRProcessingINHandoverHIWorkstationHDDecisionQCQuality Check

    Speed 1× · 0 processes running · 0 runs completed

    Simulated runs of this panel, not a measurement of a live system. The time spent at each station is the total duration of the process type, spread evenly across its stations. Which type starts next depends on its configured share. A new process every 2 s at 1× speed.

    The diagram is shown from a screen width of 640 px on. 13 nodes in 5 layers, 13 connections, 4 processing paths. To the parts list

    Legend

    A building block with at least one connection in use, shown with a solid outline.
    A building block with no connection in use, shown with a dashed outline (4: Customer Portal, Internal Systems, Partner Access, Decision).
    A connection, derived from the station sequence of the processing paths. Connections to external services are not recorded in the data and are therefore not drawn.
    L
    Column header for the layer a building block sits in, with the count of how many sit in it next to it. The data does not name the layers.
    ALT
    A process in progress, a dot carrying the short code of its type. It follows exactly the station sequence of its path.
    Station
    A working station. Its outline stands out and progress runs along its bottom edge. Each station handles one process at a time. The dots next to it are waiting.
    Speed
    Speed controls only the playback rate (0.5× / 1× / 2×). It changes nothing about the paths and dwell times, which stay exactly as recorded in the data.
    Sources
    Speed and display mode change only the simulation, not the process paths shown.

    Public BKS agent

    Talk to a running BKS agent

    The public agent answers questions about BKS-Lab over the A2A protocol. It works exclusively with public information and knows no customer or project data.

    Your input is transmitted to the agent. Please do not enter confidential information.

    Open the public BKS agent →

    From structural diagram to live demo
    The diagram and simulation explain the principle of modular agent systems. The linked BKS agent is a running service and answers general questions based on public information.

    Back to the Labor overview →

    Which task should run as its own specialist?

    We assess where clearly scoped agent roles can complement existing workflows and where a classic integration remains the better fit.

    Schedule a conversation