Agents & Runs
Manage your agents from the Voice Agents list, configure per-agent settings, and inspect run history and individual call details.
The Voice Agents page (/workflow) is where you list, organize, and manage your agents. This page covers the list view, per-agent settings, and run history and detail views.
This page does not cover the drag-and-drop workflow builder itself — that's documented in Voice Agent Builder and the rest of that section. Anywhere the API or backend says "workflow," think "agent" — see Workflows & Agents.
The agents list
Under Active Agents, your non-archived agents are shown in a table with columns ID, Agent Name, Created At, Total Runs, and Actions.
If you've created folders, agents are grouped into collapsible sections — one per folder, plus an Uncategorized section for agents that aren't in a folder. If you haven't created any folders, agents render as a single flat table instead.
Each folder section has Rename and Delete actions. Deleting a folder does not delete its agents — they move to Uncategorized.
An Archived section appears only if you have archived agents, with a restore action for each.
Row actions
| Action | What it does |
|---|---|
| Edit (pencil icon) | Opens the agent in the workflow builder |
| Move | Moves the agent to a different folder, or to Uncategorized |
| Archive / Restore | Archives an active agent, or restores an archived one |
Creating an agent
The Create Agent button offers three starting points:
- Use Agent Builder — describe your use case in natural language and Talkr generates an initial agent for you, using a Talkr-owned LLM key (not your organization's own model configuration). Limited to 10 generations per organization per day.
- Blank Canvas — starts from a single Start Call node so you can build the graph yourself. No daily limit — use this if you've hit the generation cap or prefer building by hand.
- Use a Template — pick from a small set of pre-built agents for common use cases (see Starter Templates for the full list) and Talkr copies it into a new agent you can edit freely. Two templates additionally include regulated-industry safety behavior out of the box — emergency triage and a no-medical-advice boundary for healthcare scheduling, and an FDCPA-aware disclosure/dispute/cease-contact flow for debt collections. A faster starting point than a blank canvas when your use case matches one of the templates.
You can also Upload Agent Definition to create an agent from a previously exported workflow JSON file.
All entry points land you in the workflow builder — see Voice Agent Builder for what to do from there.
Per-agent settings
Open an agent and go to Settings to configure everything that isn't part of the conversation graph itself. The settings page is organized into sections, navigable from a sticky side panel:
| Section | What it configures |
|---|---|
| General | Agent name; ambient background noise (built-in or a custom upload, with a volume control); turn detection strategy (transcription-based or Smart Turn Analyzer) and its timeout; interruption strategy; enhanced timestamped transcript; context compaction; max call duration and max user idle timeout; optional external PBX field mappings |
| Model Overrides | Override the organization's model configuration for this agent only — see Model Configurations |
| Template Variables | Key/value defaults available to this agent as {{variable_name}} — see Context & Variables |
| Dictionary | Custom terms/pronunciations for this agent (may incur extra provider cost) |
| Voicemail Detection | Enable voicemail detection, choose whether to reuse the agent's LLM or configure a dedicated one, customize the classification prompt, and set the speech-cutoff timing |
| Language | Primary and additional languages, sourced from your organization's configured voice profile |
| Recordings | Informational — recordings are managed organization-wide; links to Recordings |
| Add to Website | Embed this agent as a web widget — see Add to Website |
| Report | Download a CSV of this agent's runs, optionally filtered to a date/time range |
| Agent UUID | The agent's stable UUID, used in agent-stream URLs and other integrations where a numeric ID isn't portable — see Agent UUID |
Each section shows an indicator when it has unsaved changes, and each section saves independently.
Run history
Open an agent and go to Runs (/workflow/{id}/runs) to see every call handled by that agent. The table shows ID, Status (Completed / In Progress), Created At, Call Type (channel and direction, e.g. inbound phone call), Duration, Disposition, QA (a Flagged badge if the call crossed your QA Alerts threshold), and row actions to preview media or open the run in a new tab.
If any run on this agent has a QA node result, a QA Summary card appears above the table: the average quality score and the most common tags across every QA-scored call for this agent, so you can spot recurring issues without opening runs one at a time.
Use the filter builder to narrow the list by:
- Date and Time Range — with quick presets, up to 30 days
- Disposition Code — multi-select; options come from this agent's configured disposition codes if set, otherwise a default set
- Call Duration — a range in seconds, with quick presets
- Completion Status — Completed, In Progress, or All
- Token Usage — a numeric range
- QA Status — Flagged, Not Flagged, or All
- Search Transcripts — find calls by what was actually said, not just metadata (see below)
Results are paginated at 50 per page, and you can sort by Duration.
Searching transcripts
Type a word or phrase into Search Transcripts and the run list filters down to calls where either the caller or the agent said something matching it — the same field you'd use to look for "refund," a product name, or anything else that came up on a call. A Match column appears showing the exact line that matched, and results are ordered by how well they match rather than by date.
Check Smart search to widen the results beyond exact wording — it also finds calls that are about the same thing even if they used different words (e.g. searching "cancel my plan" can also surface a call where someone said "I want to stop my subscription"). Plain search is instant; Smart search takes a little longer since it has to understand the meaning of your query, not just match the text.
Search only covers calls indexed after this feature shipped. Older calls are backfilled automatically in the background — if a call you know exists isn't showing up in search yet, give it a little time.
For run history across all your agents at once, with per-run cost and CSV export, use Agent Runs instead.
Run detail
Clicking into a run (/workflow/{id}/run/{runId}) opens the full record of that call:
- Header — agent name, Agent ID and Run ID (with copy buttons), call time, and a Customize Agent link back to the builder. If the run produced a trace, a View Trace link is shown; if your organization has Grafana connected, a View in Grafana link is also shown, scoped to this call's own time window.
- Downloads — preview, download transcript, and download recording (voice runs). The preview dialog also lets you leave a Good / Bad / Flag verdict and an optional note on the call — a lightweight way to record human review alongside (or instead of) the automated QA node's output. Verdicts are visible to anyone else who opens the same run's preview.
- Run Metrics — duration, user turns, bot turns, tool calls, and nodes visited.
- Turn Timeline — for voice runs, a per-turn breakdown: a stacked latency bar (speech-to-text, model, and text-to-speech time) for each turn so a slow step is obvious at a glance, the sequence of nodes the call actually passed through, and any tool calls made along the way.
- Split Tracks — when separate user and bot audio tracks are available, a synced waveform view lets you play, toggle, and download each track independently.
- Initial Context and Gathered Context — the data passed in when the call started, and the structured data the agent extracted during the call, both as inspectable JSON.
- QA Results — a structured per-node breakdown (score, sentiment, tags with reasons, summary) if a QA node ran on this agent, plus a Needs Attention banner if the call crossed your organization's QA Alerts threshold.
- Conversation — a turn-by-turn replay of the call in the right-hand panel.
Text-chat sessions (agents embedded as chat rather than voice) show the same context and metrics, without the audio-specific sections.
If a run hasn't completed yet and isn't a text-chat session, the page shows a placeholder instead of run details.