Talk to your custom agents

Highlight

Talk to your custom agents. Custom agents have always worked for you in the background, a schedule fires, the agent runs, a report lands. This week, conversation became the way to work with them in both directions: chat with any agent you've built, and build new ones by describing them.

Open a live conversation with an agent and it answers as the specialist you briefed, your instructions, your data scope, your rules about what it can touch:

• Two ways in. Select New chat on an agent's page, or type @ in the message box and pick the agent. Either way the conversation is bound to that agent from the first message and labeled with its name.
• It knows its own setup. Ask an agent what it watches, when it runs, or what it's allowed to change, and it answers from its live configuration, useful for auditing an agent you set up months ago.
• Its execution mode holds in chat. A report-only agent's conversation stays advisory, it investigates and tells you what it *would* do, but can't make changes, no matter how the request is phrased. An auto-execute agent can act mid-conversation, with sensitive actions pausing for your approval as always.

Read more in Chat with a custom agent.

Building an agent works the same way, describe what you want, answer a couple of questions, and Nash assembles it for you:

• Describe it, and Nash builds it. Tell the agent what you want in plain language, "watch for late pickups each morning and flag anything trending the wrong way", and it asks a short set of questions, each with a recommended default, then assembles the agent so you don't have to fill out the builder by hand.
• Turn a conversation into a standing report. When a chat session has done something worth repeating, choose Schedule as a report from the session menu. Nash turns it into a scheduled agent, drafting the name, cadence, time zone, and recipients into a card you edit and confirm.
• Author skills by describing them. Describe a playbook you want your agents to follow and Nash drafts the skill, its name, its when-to-use trigger, and the procedure, then shows a confirmation card before saving, and a field-by-field diff when you're updating one.

The builder you fill in by hand got shorter too, see below.

New & improved

Nash Agent

• The agent builder is now organized into three steps, Basics, Execution, and Output, with less-common settings tucked into an Advanced section, so the common path is shorter.
• A new Only notify on certain results control lets an agent notify you only when a result matches a condition you set, for example, only when a check comes back failing, instead of on every run.
• Custom agents now run on a model Nash picks for the job, so there's no model to choose, the model selector is gone from the builder.
• Agent descriptions are written for you from the agent's name and instructions, so there's no description to fill in.
• Notification templates can reference {{agent_response}} to drop the agent's written response into the message.
• A quick-nav rail in agent chat lets you jump between your messages in a long conversation.
• The Nash Agent voice call can now tell a caller what canceling a delivery would cost, and what was charged on one that's already canceled.

Workflows

• The Driver Stacking alert can be set to fire only when the same driver is booked across *different* providers, so intentional single-provider batching no longer triggers it. *(Opt-in.)*

Orchestration

• Nash can hold a delivery whose pickup store is already past closing time, rather than reassigning it to a provider that can't complete the pickup. *(Opt-in.)*

Integrations

• Added Australia Post as a delivery provider in Australia for merchant-allocated shipments, Nash creates and manifests the shipment and forwards your article IDs, barcodes, packaging type, and email tracking. *(Opt-in.)*

Portal

• Route activities now show forms submitted at the route level, not only per-job submissions.

Fleet

• The driver app groups the package list by stop on pickup and return runs, items sit under their destination stop, with the stop number, recipient, and a count, so drivers load and hand off stop by stop.
• Drivers can complete a custom check-in form your organization has configured, rendered in the app at check-in. *(Opt-in.)*

Fixes

Nash Agent

• Agent notifications no longer repeat the report body, emails and Slack posts show it once, and markdown tables render as real tables instead of raw text.
• Agents reason from the correct current time, so elapsed-time and SLA calculations come out right.
• You can approve, reject, or dismiss a pending action straight from the agent's Needs review card, so the count clears even when there's no chat thread to open.
• A custom agent's detail header shows a single On/Off switch that matches whether the agent is actually paused, instead of an "Off" label next to a Pause button.
• A notification trigger attached to a custom agent now turns on automatically, so the agent starts running as soon as you connect it.

Workflows

• The Job Tags workflow filter matches the tags you set on your orders, so conditions built on it fire as expected.

Integrations

• JustEat orders in the EU, including Spain, dispatch reliably, Nash no longer sends fields the market rejects and always includes the required order identifier.
• Uber pick-pack-and-pay grocery orders record a pickup update when the courier finishes shopping, so that status and its webhook arrive on time.

Portal

• The Invoices card no longer shows a "Billed monthly" label for organizations on other billing cycles.
• Text that was unreadable in light mode now renders correctly, provider support details, tooltips, and status and error messages.
• Opening an optimization run or route map no longer fails when a stop has invalid coordinates, and meal and break stops are labeled as breaks rather than dropoffs.

Fleet

• Map links from the driver app open directly in driving-directions mode instead of dropping a plain map pin.