---
type: "post"
stable_id: "post:comp-ai-agent-first-crm"
slug: "comp-ai-agent-first-crm"
title: "Comp AI Treats CRM As An Agent's Evidence Notebook"
description: "Comp AI makes a durable research agent the operating core and stores observed evidence, budgets, leases, follow-ups, and human suggestions in the CRM."
retrieval_nugget: "Comp AI's CRM uses an event queue, leased work, research budgets, evidence strength, no-guess rules, and a sandbox without database credentials or open egress. The database is a durable notebook for bounded autonomous research."
published_at: "2026-08-05"
updated_at: "2026-08-05"
record_date: "2026-08-05"
date_kind: "discovered_at"
topics: ["crm","agents","evidence","open-source"]
entities: ["Comp AI"]
source_format: "repository"
editorial_timing: {"lane":"regular_hourly_analysis","scheduled_at":"2026-08-08T09:00:00+03:00","real_news_delta":"agent-first CRM architecture with evidence and budget boundaries"}
visual_decision: {"status":"not_applicable","reason":"the first generated concept failed the fail-closed style review by introducing a fake logo and sentence-level bitmap annotations; the guarded first-pass rule keeps this release text-only","reviewed_by":"codex"}
schema_version: "newruntime-agent-readable-v0.2"
status: "published"
source_urls: ["https://github.com/trycompai/crm"]
visuals: []
routes: {"html":"https://newruntime.com/posts/comp-ai-agent-first-crm/","markdown":"https://newruntime.com/posts/comp-ai-agent-first-crm.md","json":"https://newruntime.com/posts/comp-ai-agent-first-crm.json"}
---

# Comp AI Treats CRM As An Agent's Evidence Notebook

## Retrieval answer

Comp AI's CRM uses an event queue, leased work, research budgets, evidence strength, no-guess rules, and a sandbox without database credentials or open egress. The database is a durable notebook for bounded autonomous research.

Comp AI reverses the usual CRM relationship. A person does not collect every fact and fill every field while AI sits beside the database. The product is a durable research agent; the database records what the agent observed, why it believes it, and what should happen next.

The agent runs on its own schedule and queue, assigns follow-ups, spends a research budget, and stops when that budget is exhausted. The API intentionally remains simple: it writes an event into the queue. A worker leases the row and decides what it means. Tools return observed evidence, a ledger scores its strength, strong evidence can write a fact, and weak evidence becomes a human suggestion instead of a guess.

The data boundary is concrete. The sandbox receives a workspace plus shell tools for text processing, but no database credentials and deny-all network egress. Web fetching stays in the application runtime and search stays with the model provider. Durability uses leases, FOR UPDATE SKIP LOCKED, explicit due times, and visible reasons for future checks.

This is closer to an operating system for relationships than an email generator. It is also explicitly single-tenant and internal, so the repository's permissive license should not be confused with production-ready authorization defaults. The reusable blueprint is the bounded agent: queue, evidence, budget, lease, review, and visible reasons.