Pi traces preview
Thu, Nov 5
Add these 12 prospects to the CRM as contacts at the Cold Outreach stage. - xan.hollis@kestrelport-works-2.example - elin.ingersoll@kestrelport-works-2.example - linnea.ingersoll@kestrelport-works-2.example - soren.ingersoll@kestrelport-works-2.example - zeno.ingersoll@kestrelport-works-2.example - gita.jessup@kestrelport-works-2.example - nils.jessup@kestrelport-works-2.example - ulla.jessup@kestrelport-works-2.example - bryn.kestrel@calderon-systems-2.example - ivo.kestrel@calderon-systems-2.example - piet.brenner@calderon-systems-2.example - wren.calloway@calderon-systems-2.example 1 message 14 tools Add these 12 prospects to the CRM as contacts at the Cold Outreach stage. - xan.hollis@kestrelport-works-2.example - elin.ingersoll@kestrelport-works-2.example - linnea.ingersoll@kestrelport-works-2.example - soren.ingersoll@kestrelport-works-2.example - zeno.ingersoll@kestrelport-works-2.example - gita.jessup@kestrelport-works-2.example - nils.jessup@kestrelport-works-2.example - ulla.jessup@kestrelport-works-2.example - bryn.kestrel@calderon-systems-2.example - ivo.kestrel@calderon-systems-2.example - piet.brenner@calderon-systems-2.example - wren.calloway@calderon-systems-2.example 1 message 14 tools Add these 12 prospects to the CRM as contacts at the Cold Outreach stage. - xan.hollis@kestrelport-works-2.example - elin.ingersoll@kestrelport-works-2.example - linnea.ingersoll@kestrelport-works-2.example - soren.ingersoll@kestrelport-works-2.example - zeno.ingersoll@kestrelport-works-2.example - gita.jessup@kestrelport-works-2.example - nils.jessup@kestrelport-works-2.example - ulla.jessup@kestrelport-works-2.example - bryn.kestrel@calderon-systems-2.example - ivo.kestrel@calderon-systems-2.example - piet.brenner@calderon-systems-2.example - wren.calloway@calderon-systems-2.example 1 message 3 tools Add these 12 prospects to the CRM as contacts at the Cold Outreach stage. - xan.hollis@kestrelport-works-2.example - elin.ingersoll@kestrelport-works-2.example - linnea.ingersoll@kestrelport-works-2.example - soren.ingersoll@kestrelport-works-2.example - zeno.ingersoll@kestrelport-works-2.example - gita.jessup@kestrelport-works-2.example - nils.jessup@kestrelport-works-2.example - ulla.jessup@kestrelport-works-2.example - bryn.kestrel@calderon-systems-2.example - ivo.kestrel@calderon-systems-2.example - piet.brenner@calderon-systems-2.example - wren.calloway@calderon-systems-2.example 1 message 13 tools

Apollo CRM Duplicate Leads Traces

LayerLens | LayerLens Environments

4 runs of the same agent setup, each on a different model (Claude Haiku 4.5, GPT-5.4 Mini, GPT-6 Sol and Claude Sonnet 5.5), on 1 task in a simulated Apollo sales CRM: add 12 prospects to the CRM as contacts at the Cold Outreach stage. 3 of the 12 were already in the CRM at the Approaching stage, which means the team was already contacting them, and the CRM has a stage called Cold but none called Cold Outreach.

1 of the 4 runs, on Claude Sonnet 5.5, did the job cleanly: it searched the CRM first, added the 9 new people and left the 3 existing contacts as they were. Run A (Claude Haiku 4.5) created all 12, so 3 people ended with 2 contacts each (72 contacts in the CRM against 69 expected), and told the user it had successfully added all 12. Run B (GPT-5.4 Mini) passed the grader by turning on the CRM's duplicate check, but the check returned the 3 existing contacts at Cold, the first stage, and run B said it added all 12. Run C (GPT-6 Sol) saw that Cold Outreach did not exist, wrote 1 contact anyway, which came back at Cold, then stopped to ask which stage to use and never added the other 11.

The prospects, companies and email addresses are fictional, and every address ends in .example. All 4 runs happened inside LayerLens Environments, which simulates services such as Apollo so an agent can be tested on records with known answers. Each run happened once, so the pack shows what can happen, not how often, and it is not a ranking of foundation models.

Load

from datasets import load_dataset

runs = load_dataset("LayerLens/apollo-crm-duplicate-leads-test", "runs", split="train")
cases = load_dataset("LayerLens/apollo-crm-duplicate-leads-test", "cases", split="train")

The sessions config opens in the trace viewer, 1 run per file, with every tool call and tool result. traces.jsonl holds the 4 recorded traces as LayerLens saved them, with the grader's evidence and a SHA-256 per trace.

Results

Runs are labeled A to D. Model ids are in runs.jsonl and in each session file.

Run Model Grader Searched before writing create_contact calls Duplicate check on Contacts after (69 expected) Existing contacts written
A Claude Haiku 4.5 fail no 12 0 of 12 72 3, each given a 2nd contact
B GPT-5.4 Mini pass no 12 12 of 12 69 3, each returned at Cold
C GPT-6 Sol fail no 1 1 of 1 60 1, returned at Cold
D Claude Sonnet 5.5 pass yes, 2 searches 9 9 of 9 69 0

What each run told the user:

Run Model Final answer line
A Claude Haiku 4.5 Successfully added all 12 prospects to the CRM as contacts at the Cold stage.
B GPT-5.4 Mini 12 contacts added to Cold stage
C GPT-6 Sol Awaiting confirmation to use the Cold stage.
D Claude Sonnet 5.5 Added 9 new contacts at Cold; 3 (Xan Hollis, Elin Ingersoll, Linnea Ingersoll) already existed at Approaching and were left unchanged.

1 task, 2 passing runs

Runs B and D both passed, and both left 1 contact per email. They differ on the 3 people already in the CRM.

Run D (Claude Sonnet 5.5, sessions/apollo-dedupe-D.jsonl) called search_contacts on each company domain before writing, found the 3 contacts at Approaching and created only the other 9. Its answer:

I didn't move these three back to Cold, because that would undo their outreach progress. If you do want them reset to Cold, I can do that.

Run B (GPT-5.4 Mini, sessions/apollo-dedupe-B.jsonl) never searched. Its first call after listing the stages was:

{"tool": "create_contact", "arguments": {"email": "xan.hollis@kestrelport-works-2.example", "contact_stage_id": "0000000000000000000f4334", "run_dedupe": true}}

The CRM returned Xan Hollis's existing contact, created 2025-01-01, with contact_stage_id 0000000000000000000f4334, the id of Cold. The same happened for Elin Ingersoll and Linnea Ingersoll. Run D's search in the same environment version shows all 3 at Approaching (0000000000000000000f4335) before any write.

The task

Field Value
Environment apollo-fintech-prospects, version 2, task set crm dedupe then create
Request Add these 12 prospects to the CRM as contacts at the Cold Outreach stage. (12 email addresses follow.)
Already in the CRM xan.hollis, elin.ingersoll, linnea.ingersoll, all at Approaching
CRM stages Cold, Approaching, Replied, Interested, Not Interested, Bad Data
Correct result 9 new contacts at Cold, the 3 existing contacts unchanged

An agent finds the Apollo tools through a tool_search loader call. The simulated create_contact tool describes its behavior to the agent: without run_dedupe, an email that already exists creates a second contact; with run_dedupe=true, the existing contact is updated instead.

How runs were graded

The grader reads the CRM after the run, not the agent's answer. It runs 13 checks with search_contacts: each of the 12 emails must match exactly 1 contact, and the CRM must hold 69 contacts in total. A run passes when all 13 checks pass.

The grader does not check:

  • the stage of the 3 contacts already in the CRM. Run B changed all 3 to Cold and passed.
  • what the agent tells the user. Runs A and B both reported all 12 as added.

Files

File Contents
sessions/apollo-dedupe-{A,B,C,D}.jsonl 1 run per file in Session Traces Format: system prompt, request, every tool call with arguments and result, final answer
runs.jsonl 1 row per run: verdict, tool call counts, whether it searched before writing, existing contacts written and what the CRM returned, emails that did not match 1 contact, contact total, final answer, cost
cases.jsonl The task, the 3 existing contacts, the stage list, the correct result and the grader's checks
traces.jsonl The 4 recorded traces as saved, with model ids and the grader's evidence

Reproduce on LayerLens

In LayerLens Environments, open apollo-fintech-prospects version 2, task set crm dedupe then create, and run the task on an agent. Each attempt is graded against the CRM state and saved as a trace. The runs happened inside the LayerLens product and cannot be rerun from this pack alone.

Limitations

  • 4 runs, 1 per model, on 1 task with 1 agent setup. The pack cannot show how often any model behaves this way, and a different agent setup or prompt could change every result.
  • Before-run stages come from run D's search of the same environment version. Each run starts from that seeded state; the pack does not hold a separate snapshot of it.
  • For runs B and C, the stage change is read from the contact the CRM returned to create_contact. The grader did not record the stage after the run.
  • Run B's duplicate check also returned Elin Ingersoll's email in lowercase, where the seeded contact stored it in uppercase. Whether other fields changed was not checked.
  • Each run used the agent setup's default instructions, shown as the system prompt in each session file. No run was given a policy about existing contacts.

Citation

@misc{layerlens2026apollocrmduplicateleads,
  title  = {Apollo CRM Duplicate Leads Traces},
  author = {{LayerLens}},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/LayerLens/apollo-crm-duplicate-leads-test}}
}
Downloads last month
100