Podmod runs six agents during a recording. Each one watches the transcript as it forms and surfaces a different kind of card in the panel beside your conversation. They are silent, visual only, and toggled individually from the Agents menu in the control bar before or during a take.
Running all six on every episode is usually the wrong call. A news recap and a comedy interview need different things on screen. This guide covers what each agent does, what it actually outputs, and which combinations suit which formats.
The Short Version
| Agent | What it does |
|---|---|
| Research Agent | Finds web links, articles, and reference material while you talk |
| Fact Checker | Checks specific claims, stats, dates, and attributions |
| Media Finder | Surfaces images and videos related to what you said |
| Producer Coach | Follow-up questions, fresh angles, devil's-advocate notes |
| Prep Desk | Pulls from your uploaded notes, briefs, and session files |
| Rundown | Auto-checks your planned questions as the conversation covers them |
Research Agent
What it does. Listens for topics and returns web sources on them: articles, reference pages, and links relevant to what you just said.
What a card looks like. A topic heading followed by two or three sources with publication names. Mention a new model release on a tech show and you get the announcement page plus the coverage of it.
When to run it. Any show where the conversation ranges across topics you have not pre-researched. News, tech, culture, anything reactive.
When to leave it off. Tightly scripted shows, or episodes where you have loaded everything you need into session context and want Prep Desk to be the source instead. Two agents returning material on the same line is noise.
Fact Checker
What it does. Watches for checkable assertions and returns a verdict with a source. Specifically: claims, statistics, dates, and attributions.
What a card looks like. A verdict label, the claim, the correction, and the source:
Fact Checker Likely incorrect "LeBron has more NBA championships than Jordan." Jordan has 6 titles; LeBron has 4. Basketball-Reference, NBA Finals history
When to run it. Anything with numbers, dates, or attributed claims in it. News, sports, finance, history, health, politics. Also any interview where a guest will state figures about their own company or field.
Why it matters. Trust in media is fragile and declining. YouGov's 2026 Trust in Media survey found trust dropped for most outlets measured. Research on corrections is nuanced: audiences are actually more distrusting of journalists who correct false claims than those who confirm true ones. The practical implication is not that corrections are bad. It is that a correction made live, inside the conversation, reads very differently from one issued afterward. Catching it in the take means you never have to issue the second kind.
When to leave it off. Fiction, comedy improv, and anything where nothing said is meant literally. Checking a bit is not useful.
Media Finder
What it does. Surfaces images and video clips related to what was just said.
What a card looks like. For images, a set of visual references: mention a restaurant on a food show and you get plating shots. For video, clips with titles: mention a specific play on a sports show and you get the footage.
When to run it. Video podcasts, obviously, where you may cut to the reference. But also audio shows, for two less obvious reasons. First, seeing the thing you are describing makes you describe it better. Second, the cards persist with the session, so they become your source list when you build the episode page or cut social clips.
When to leave it off. Audio-only shows where you find visual cards distracting. Some hosts do. It is a legitimate preference.
Producer Coach
What it does. Suggests follow-up questions, alternative angles, and devil's-advocate positions, tied to what the guest just said.
What a card looks like.
Producer Coach Ask what the first week without a paycheck felt like. Money anxiety lands harder than the resignation story.
When to run it. Every interview. This is the agent that most directly improves the conversation rather than the accuracy of it. Interview research consistently finds that follow-up questions are where the best material emerges, and follow-ups are the thing a prepared question list structurally cannot contain, because they depend on an answer that has not happened yet.
When to leave it off. Solo episodes, where there is no one to follow up with. Also panel shows that are already crowded with voices.
How to use it well. Read the suggestion, then ask it in your own words. A card is a prompt, not a script. Reading it verbatim sounds like reading it verbatim.
Prep Desk
What it does. Draws on the material you uploaded to the session and surfaces it when the conversation reaches it. Notes, files, URLs, and guest bios loaded through the Context panel.
What a card looks like.
Prep Desk Guest bio: sold prior SaaS for $12M in 2021 after two failed pivots. Ask what they kept from those failures. In response to: "we pivoted hard after the first product flopped"
When to run it. Any episode you prepared for. Which should be most of them.
Why it is underrated. The other agents bring you information you did not have. Prep Desk brings you information you already found and could not recall in the moment. For most hosts, the second category is larger.
Requirement. It needs material. Prep Desk with an empty Context panel does nothing. Load the session first.
Rundown
What it does. Watches the conversation against your planned question list and checks items off as they are genuinely covered.
What a card looks like. Your list with a live count, checked items struck through, and auto-handled items marked.
When to run it. Structured interviews, solo episodes with an outline, and any show where you routinely run out of time before the closing segment.
When to leave it off. Free-form conversation shows with no planned structure. There is nothing to check.
Combinations by Format
Guest interview. Prep Desk, Producer Coach, Rundown, Fact Checker. This is the workhorse set. Your prep surfaces, your follow-ups improve, your structure holds, and your guest's claims get checked.
News or recap show. Research Agent, Fact Checker, Rundown. You need sources fast, accuracy on figures, and a stack you can clear against runtime.
Solo commentary. Prep Desk, Research Agent, Rundown. No one to coach you into follow-ups, so replace Producer Coach with more source material and hold your outline.
Video-first show. Add Media Finder to whatever else you are running. If you are cutting to visuals, you want them found during the conversation rather than after it. Edison Research now finds YouTube is the most-used podcast service among weekly consumers, which makes visual assets part of the core deliverable rather than an extra.
Comedy or narrative. Producer Coach only, or nothing. Accuracy agents fight the format.
Practical Notes
Toggle mid-session. You can turn agents on and off during a recording. If the Fact Checker is firing on every sentence in a segment that is all opinion, switch it off for that stretch.
Fewer agents means better cards. Six agents on a fast conversation produces a panel you cannot read. Three produces a panel you can. Pick what the episode needs.
Nothing is audible. No agent makes a sound or shows anything to your guest. Cards appear on screen for the host only.
Cards persist. Everything surfaced during the take is saved with the session and appears in Session Viewer alongside the transcript, anchored to the line that triggered it. The research does not evaporate when you stop recording.
Personalization applies to all of them. Custom instructions set in the Personalize panel shape how the agents behave on your show. You can write them, or paste a link to your show and let Podmod research the channel and draft producer-style guidance for your format.
How the Agents Decide What to Respond To
Worth understanding, because it explains why cards sometimes do not appear when you expect them.
The agents work from finalized phrases in the live transcript, not from raw audio. A thought has to complete before it can be evaluated. This means there is a short lag between saying something and seeing a card about it, and it means half-finished sentences that you abandon mid-way generally do not trigger anything.
It also means the agents respond to what was actually said rather than what you meant. If a name is mispronounced badly enough that the transcript does not catch it, the Fact Checker cannot check it. This is rarely a problem in practice, but it explains the occasional miss.
The practical takeaway is simple: the agents work best on shows where people speak in reasonably complete thoughts. Rapid crosstalk produces fewer useful cards than a conversation where people finish sentences.
Start With Three
If you are setting this up for the first time, do not turn on all six. Run Prep Desk, Fact Checker, and Producer Coach on your next interview. That combination covers recall, accuracy, and conversation quality, and it produces a panel you can actually read while talking.
Add the others once you know what your show reaches for.
Try it on your next episode at app.podmod.ai.