The first AI agent is usually a product experiment. The second is where the operating problem begins.
A support agent may start on a web page. Then sales wants an agent for inbound email, and operations wants another agent to answer calls and texts. Without a shared operating model, that quickly becomes three runtimes, three deployment processes, and three different ways to answer a basic question: what version is running right now?
The open-source agent-fleet sample demonstrates a smaller operating surface. It contains three specialized agents:
-
acp-support: browser-based support for returns, shipping, and account-change requests -
acp-sales: an inbound email agent that answers questions, qualifies prospects, and suggests a next step -
acp-scheduling: a voice and SMS agent that collects a customer's name, requested job, and preferred time
Each project contains only two files: agent.json and instructions.md. There is no application server or channel-specific integration code in the sample.
Two files define each agent
The repository is deliberately repetitive:
agent-fleet/
├── acp-support/
│ ├── agent.json
│ └── instructions.md
├── acp-sales/
│ ├── agent.json
│ └── instructions.md
└── acp-scheduling/
├── agent.json
└── instructions.md
agent.json declares the agent's model, channel configuration, memory, and required secrets. instructions.md defines the behavior and business boundaries.
For example, the support instructions tell the agent to keep replies short, request an order number for returns, and avoid inventing addresses or deadlines. The scheduling instructions collect three details but explicitly prohibit promising a confirmed appointment. These operational limits live beside the configuration and can be reviewed like any other source file.
All three agents use Telnyx-hosted zai-org/GLM-5.3-Flash, so the projects do not require a separate model-provider key.
Test an agent before deploying it
Start with the support agent locally:
cd agent-fleet/acp-support
telnyx-acp validate
telnyx-acp dev
validate checks the configuration and reports the field and suggested correction when it finds a problem. dev starts the local development loop on port 8787.
In a second terminal, open a local chat session:
cd agent-fleet/acp-support
telnyx-acp chat --local
Try asking:
Where do I send a return?
The expected behavior is not to fabricate a return address. The agent asks for the order number, confirms it, and explains the next step established in its instructions.
This local loop is useful because prompt and policy changes can be tested before any public channel is attached.
Deploy the fleet
The scheduling agent needs a voice- and SMS-capable number. Use placeholders rather than committing a real number:
cd ../acp-scheduling
telnyx-acp numbers
telnyx-acp add phone <VOICE_AND_SMS_NUMBER>
telnyx-acp deploy --dry-run
telnyx-acp deploy
The dry run previews the plan without changing account resources. The actual deployment builds the agent and wires the attached number to its voice and messaging resources.
Deploy the other two projects with the same command:
cd ../acp-support
telnyx-acp deploy
cd ../acp-sales
telnyx-acp deploy
The support deployment prints a public web-chat location at <mount>/web. The sales deployment prints its inbound email address. Initial builds can take about ten minutes, so pre-deploying before a live demonstration is sensible.
One operating workflow across different channels
The value of the fleet is not that all three agents share the same job. They do not. The value is that they share the same operating commands.
telnyx-acp status
telnyx-acp logs --since 10m
telnyx-acp versions <agent-name>
telnyx-acp deploy
-
statusreports the active function, revision, attached channel resources, and webhook health. -
logsshows recent runtime activity and failures. -
versionslists immutable deployment revisions with their Git short SHAs. -
deploycreates a new revision after a configuration or instruction change.
Web and email conversations also support traces:
telnyx-acp traces email-<thread-id>
telnyx-acp traces email-<thread-id> <run-id>
The first command lists runs for the conversation. The second walks through a specific run's model and memory steps.
Phone calls are different in the current beta: call traces are not available, so use telnyx-acp logs to inspect call-side behavior.
Memory stays scoped to the conversation
Each agent declares an Agent Memory profile using {{conversation}} as its key. The actual key depends on the channel:
- the browser conversation ID for web chat
-
email-<thread-id>for email - the caller's number for phone and SMS
That gives each conversation durable context without requiring the sample to implement its own session database.
It is important not to overstate this behavior: these channel identifiers are different. The sample does not automatically prove that a browser visitor, an email sender, and a caller are the same person. Sharing memory across those identities would require an explicit identity-resolution strategy.
Revise an agent without changing application code
Suppose the scheduling agent needs a warmer greeting. Edit its instructions and deploy a new revision:
cd acp-scheduling
$EDITOR instructions.md
telnyx-acp deploy
telnyx-acp versions acp-scheduling
This is the core Agent Control Plane pattern: behavior is configuration, deployments are versioned, and operations use the same commands in every agent directory.
Current beta limitations
The example documents several boundaries worth understanding before production use:
- There is no fleet-wide
telnyx-acp listcommand. Runstatusinside each project directory. - There is no rollback command for CLI-deployed agents. Check out the earlier configuration and deploy it again.
- Phone calls expose logs but not traces.
- A personal number may require 10DLC registration for application-to-person SMS. The sample's CLI-based Telnyx-to-Telnyx test path avoids that requirement for a controlled demo.
The configuration-only approach also does not remove the need for production policy review, monitoring, access controls, or channel compliance.
The reusable pattern
Agent Fleet is less about three specific departments and more about specialization without operational fragmentation.
You could replace the sample instructions with agents for property maintenance, healthcare intake, field-service dispatch, or customer onboarding. Each agent can keep a narrow job and channel-specific behavior while the team retains one validation, deployment, logging, tracing, and versioning workflow.
Start with one project, test it locally, deploy it, and learn the operating commands. The second and third agents then become additional configurations rather than entirely new integration stacks.
Top comments (0)