Editing with Tandy
Ask Tandy to change an existing voice agent's settings in plain language, and apply the changes from a single confirmation card.
Tandy can change a voice agent you already have. Describe what you want different and it reads the agent, works out which settings that means, and offers one confirmation card. The agent changes when you press Apply, and not before.
This is the general editor. For hardening an agent against what a caller could abuse, use Security Review instead: it runs a checklist first and can only move a setting toward the safer value.
Asking for a change
Name the agent and say what you want:
Change Riley's greeting to "Thanks for calling Northside Dental, this is Riley."Name the agent however you like: its name, its id, or a description of it ("the agent on the support line"). If more than one agent matches, Tandy asks which you meant rather than guessing.
You can ask for several changes at once, and you can describe the outcome rather than the setting:
Riley should stop taking keypad input, cap calls at ten minutes, and hang up on a silent lineTandy reads the agent's current configuration before it proposes anything. That read is what keeps a change from clobbering the rest of a settings group, so a request it cannot read the agent for is refused rather than guessed at.
What you can change
Any setting on the agent, across every configuration page:
| Area | Examples |
|---|---|
| Persona | Name, system prompt, greeting, dispositions, variables |
| Voice and speech | Voice, language, engine, speaking rate, barge-in, end-of-turn timing |
| Returning callers | Caller memory, the returning-caller greeting, reconnect after a drop |
| Timeouts & Limits | Idle behavior, call-duration limits, retry and timeout tuning |
| Brain | Model, reasoning, conversation-history strategy |
| Tools | Workflow tools, and the knowledge base attached |
| Built-in tools | Keypad capture, recording control, transfer destinations |
| Skills | Which platform skills are on, and your own |
| Safety | PII redaction, steering rules |
| Post-call analysis | Whether it runs, and what it extracts |
Unlike a security review, this card is not limited to the safer direction. If you ask for memory sharing to be turned on, that is what the card offers. Read the card before pressing Apply: it lists every setting that will change and what it will become.
Adding a tool
Tandy can add a tool to an agent that already exists, and it resolves what the tool calls by looking it up rather than guessing:
- A workflow tool calls a workflow you already have. Name the workflow and Tandy finds it. If it doesn't exist yet, ask Tandy to create the workflow first, then ask for the tool on a later turn.
- An integration tool calls one of your REST or SOAP integrations. Tandy reads your integrations by name and needs you to say which endpoint on it to call.
- An MCP tool calls one of your MCP servers. Tandy needs only the server. If that integration has no tools enabled, Tandy says so instead of adding a tool the agent can call nothing with.
You never give Tandy a key or a password for any of these. The integration holds its own credential and the platform attaches it when the agent calls out.
Tools are an edit, not part of creating an agent. A brand-new agent Tandy builds for you can carry a workflow tool and nothing else, so ask for the integration or MCP tool once the agent exists.
What Tandy won't change
- A credential. Tandy is never told the value of an agent variable, so it cannot rotate one. Paste the new value in the agent editor yourself.
- A native speech-to-speech model. Ask in the agent editor instead.
Variables are merged, not replaced
Tandy never sees the values of your agent variables. It is told each variable's name, how long the value is, and whether the value looks like a credential. So a change to variables merges by name instead of rewriting the list:
- A variable you asked about is changed.
- A variable you did not mention keeps its value, untouched.
- Asking to remove one removes it.
That is why "give api_base a clearer description" leaves api_key alone rather than
blanking it.
Applying the changes
One card, listing every change. Nothing happens until you press Apply, and nothing is applied that is not on the card.
The card is all-or-nothing: there is no per-change checkbox. If you only want some of what Tandy proposed, decline the card and make those changes in the agent editor.
A long value is described rather than printed. A rewritten system prompt shows as
Rewritten with Tandy's reason, and a replaced list shows how many entries it will hold,
because a wall of text on a confirmation card looks reviewable and is not. Ask Tandy to
read a value back to you before applying if you want to see it.
Troubleshooting
Tandy says it can't change an agent. Changing one needs to read your agents first, and that read wasn't available for this conversation. Start a new chat and try again.
"This agent was changed while these updates were being drafted." Someone saved the agent between Tandy reading it and you pressing Apply, so nothing was applied. Applying would have reverted their change. Ask again to work from the current configuration.
A change was refused when I pressed Apply. The setting has a limit Tandy proposed a value outside of, and the error names the field. Tandy can correct it and offer a new card. The one refusal it cannot correct is Redact caller PII on an agent whose model isn't hosted on Bedrock: redaction is delivered by a Bedrock guardrail, so changing model vendor is the only way to get it on that agent.
Tandy changed more than I asked. Asking for a whole settings group replaces that group, so fields you didn't name go back to their defaults. Name the individual settings instead, and check the card's rows before applying.
I wanted a new agent, not a change. Tandy can create one too. See Build your first agent.