That is the part every general enterprise explainer skips, and it is the part that decides whether a recruiter trusts the output. This page covers the catalog: what it is, why choosing an agent is a different problem from connecting one, what an entry has to disclose, and which agents a recruiting team actually needs.
Key takeaways
- An AI agent marketplace is a catalog of prebuilt agents you can evaluate and switch on without building them.
- For recruiting, the useful catalog entries name a single job in the hiring process rather than a general capability.
- A catalog does not add data. It tells you what an agent reads, whose method it follows, and when that data was refreshed.
- A catalog entry you can trust discloses six things, including what the agent will refuse to answer.
- Findem publishes a named lineup: Calibration, Application Boost, Screening, Scheduling, ID Verify, Veteran Sourcing, Fia and Intelligent Job Post.
- A catalogued agent still does not verify that a candidate can do the work. An assessment does.
Why is choosing an agent a different problem from connecting one?
Connection is a plumbing question with a technical answer. Selection is a judgment question with no technical answer at all. Once a connection exists, an assistant can reach your data, and the next question is which agent should act on it, using whose method, with what disclosure. No protocol answers that.
Here is the concrete version. Two agents in a catalog both offer to screen applicants for a warehouse role. One reads the resume text. The other reads the resume, the application answers and the assessment result, refuses to score on anything outside the calibrated requirement list, and returns a reason for every recommendation. Both connect identically. Only one of them will survive a conversation with your legal team.
What this page does not cover. The protocol underneath a catalog, how MCP connections are authorized and what a security review asks about them, is covered in MCP for recruiting. For what agents do hour to hour in a recruiter's week, see AI agents in recruiting.
What does a library catalog actually tell you?
A library catalog does not add a single book to the shelves. It tells you what is on the shelf, which edition it is, when it arrived, and whether it has been superseded. Take the catalog away and the books are still there, but nobody can tell what they are looking at or whether it is current. That is the whole function, and it is why a catalog is worth more than the sum of its listings.
An AI agent marketplace for recruiting has to carry the equivalent metadata about people data, which comes down to three facts per entry:
- Where the record came from, so a recruiter knows whether a claim is self reported, inferred or verified.
- How the role was labeled, so the definition of a senior data engineer is written down rather than assumed.
- When the data was last refreshed, because a candidate who changed jobs in March makes a February profile wrong.
Findem's approach to the third fact is labeling at the data layer. Findem publishes 1B+ career paths mapped, 2M+ labeled Success Signals, 200+ contributing experts and 75 to 100 Success Signals per profile on its platform page. Those labels are what a catalog entry can point at when it claims to know something about a person, and an agent reading unlabeled records has nothing equivalent to show you.
How is a catalogued agent different from automation, a chatbot and a bare connection?
The four options differ most in what you can inspect and how they fail. A rules based workflow fails loudly and predictably. A chatbot fails fluently. A bare connection fails by answering a question it should have refused. A catalogued agent is the only one of the four that tells you in advance what it reads and what it will not do.
| Comparison Criteria | Rules based automation | Chatbot | Bare MCP connection | Catalogued agent |
|---|---|---|---|---|
| What you give it | A trigger and a fixed condition | A question | A credential and a prompt | A goal and a calibrated definition of the role |
| What it does | Runs the same steps every time | Answers from what it was trained or fed | Fetches whatever the tools expose | Works a named job end to end within its stated scope |
| What it hands back | A completed action or an error | Text, usually confident | Raw or summarized records | Finished work with a reason and a source |
| What you can inspect | The rule | The transcript | The call log, if anyone enabled it | The disclosed method, the data read, the refresh date |
| How it fails | Stops, visibly | Sounds right and is wrong | Answers a question it should have declined | Returns nothing and says why, when built well |
The practical takeaway for a recruiting team is that automation and chatbots are still the right tools for parts of the process. Reminder emails should be rules. Candidate FAQ answers can be chat. Judgment work that touches a candidate record belongs with an agent whose method you can read. That distinction is worked out in more detail in this framework for judging any AI agent and in what hiring automation covers.
How do you judge whether a catalog entry can be trusted?
A catalog entry you can trust discloses six things. Ask them as literal questions, and treat a missing answer as the answer.
- What data does this agent read? Name the systems and the fields, not "your ATS data."
- Whose method does it follow? A published rubric, a hiring manager's calibrated requirements, or an undisclosed model judgment.
- What does it check before it returns anything? Completeness of the record, recency, consistency with the requisition.
- What will it refuse to answer? An agent with no refusal conditions has not been scoped.
- When was the underlying data last refreshed, and how often does that happen?
- Who reviews the output before it reaches a candidate or a hiring manager, and what can that person see?
Question four is the one that separates a real catalog entry from a marketing page. An agent that will answer anything has no boundaries, which means nobody wrote down what it is for. Refusal conditions are a design decision, and a vendor that can state them has thought about where its agent stops.
Question two carries the legal weight. If an agent's recommendation substantially assists an employment decision, the method behind it has to be defensible under the EEOC guidance on AI and Title VII, and in New York City it may trigger a bias audit obligation under the Automated Employment Decision Tools rules from NYC DCWP. "The model decided" is not a method. The NIST AI Risk Management Framework gives your team standard vocabulary for documenting the answer.
For the wider version of this question, including what to do when the answers are thin, see how to decide whether you can trust an AI recruiting tool.
Which agents does Findem publish today?
Findem publishes a named lineup on its meet the agents page, all presented as available: Calibration Agent, Application Boost Agent, Screening Agent, Scheduling Agent, ID Verify Agent, Veteran Sourcing Agents, Fia, which is a voice and chat assistant, and Intelligent Job Post, which followed Findem's acquisition of Getro. The Screening Agent and Scheduling Agent pages are published under Glider AI branding, as Glider AI Screening Agent and Glider AI Scheduling Agent, which is how the assessment and screening side of the portfolio shows up inside the lineup.
Underneath them sits Findem Studio, which exposes Findem's real time talent graph as MCP tools for agents and workflows. Studio is the access layer, not the catalog, and the distinction matters when you are comparing vendors: one is how an agent reaches labeled people data, the other is which agent you switch on. Who gets Studio access and on what terms is covered in the Findem Studio access model, and the competitive comparison sits in Findem Studio versus SeekOut, Juicebox and Gem.
Findem does not make employment decisions. Every agent's output is a recommendation a person reviews, which is the design constraint that makes the rest of this page workable.
What does a working agent lineup look like in practice?
A lineup works when each agent owns one step and a person owns the handoff between steps. Using the agents Findem publishes, the sequence across a requisition looks like this.
- Calibration Agent, at the start. The role gets a written definition before anyone sources against it. The hiring manager signs off on that definition, and nothing downstream runs on a role nobody agreed to.
- Intelligent Job Post, next. The posting reflects the calibrated role rather than last year's template.
- Veteran Sourcing Agents, where a specific talent pool is the goal. Sourcing agents work best against a definition, which is why they come after calibration.
- Application Boost Agent, on applicants already in the pipeline. Most teams have more unreviewed inbound than they admit.
- Glider AI Screening Agent. Recommendations arrive with a reason, and a recruiter reviews before anything reaches the candidate.
- Fia, for voice and chat contact, and the AI voice recruiter pattern it belongs to.
- Glider AI Scheduling Agent. Scheduling is the step where automation pays back fastest, because the judgment already happened.
- ID Verify Agent, alongside identity verification, before an assessment result counts for anything.
Two human checkpoints matter more than the rest: the hiring manager signing the calibrated role at step one, and the recruiter reviewing screening output at step five. Skip either and you have a fast pipeline with no accountable decision in it. For teams running this at scale, the volume case is covered in high volume hiring.
Who decides when an agent hands something back?
You do, in the catalog configuration, before the agent runs. Handback rules are part of the entry: at what confidence the agent stops, which fields must be present for it to proceed, and which actions always require a click. An agent that only hands back when it fails is configured wrong, because failure is the case it is least likely to notice.
Write handback into the step rather than into a policy. The recruiter reviewing screening output should see the record the agent read, the requirement it matched against, and the date that record was refreshed. If any of the three is missing, the review is theater. What happens when an AI recruiting agent gets it wrong is mostly a story about handback rules nobody set.
What does a catalog not replace?
Verification that the candidate can actually do the work. A catalog tells you where a claim came from and how fresh it is. It cannot tell you whether the person who wrote "eight years of Kubernetes" can debug a failing cluster, and no amount of labeling closes that gap.
That is a separate layer of your stack, and it is evidence rather than inference:
- A skills assessment produces a score against a defined requirement.
- Coding simulations show work under conditions that resemble the job.
- Agentic AI interviews probe the reasoning behind an answer rather than the answer alone.
- The candidate 360 view puts profile, screening and assessment evidence in one place so a reviewer sees the whole record.
A catalog also does not replace the decision. Under the Uniform Guidelines on Employee Selection Procedures, a selection procedure has to be job related, and that obligation belongs to your organization regardless of which agent produced the recommendation.
What should your team do right now?
Pick the single step where your recruiters lose the most hours to work that has no judgment in it, and read the catalog entry for the agent that covers it. Run the six questions against that entry. If the vendor cannot say what the agent reads, whose method it follows and when the data was last refreshed, you have learned something useful about the vendor, and you have spent twenty minutes.
Then run one agent on one requisition, with one named reviewer, for two weeks. A lineup of eight agents installed in a quarter is a governance problem. One agent with a reviewer who can see what it read is a working process you can extend.




