Case study · WorkZeal

Insider

AI agents that do the work, not just chat

A manager writes a job in plain words and picks a schedule. At that time an agent wakes up, researches the web, adds what it found to the CRM through MCP with that manager's rights, and reports back.

See how it works Discuss your AI project

2agents per job: one reads the web, one writes
30minutes: life of the user's token per run
5companies at most per run, checked in the database
1log row and email report per run

Find and add 5 new potential customersjob · daily at 9:00 · as Anna Reed

Claimed by the scheduler09:00:04 next run already set, so no one else takes it

Search agent: 9 candidates09:01:12 web search on · CRM tools off

4 already in the CRM, removed09:01:13 matched by name, domain or email, in code

CRM agent: 5 companies added09:02:40 web off · insert_company ×5 through MCP

Checked in the database09:02:41 report emailed to Anna · new-company rules start

RESULT: OKAutomation → Log

WorkZeal Insider · scheduled AI agents · MCP

How a job runs

Each job is one plain-language instruction with a schedule and a responsible person. The run is split between two agents with different powers, and plain code sits between them: it removes duplicates, caps the volume and checks the result in the database.

Two agents, two sets of powers. The one that reads the web never gets CRM tools, and the one with CRM tools never sees the web, so text planted on a web page can't turn into an action in the CRM.

Why two agents, not one

Search agentCan read the whole web. Can't touch the CRM. Its answer is only a list of candidates.

Code in betweenRemoves companies already in the CRM or found before, and keeps 5 at most. No model decides this.

CRM agentCan write to the CRM, as one user. Can't browse, so web pages can't give it orders.

How I built it

1Schedules and claiming

Once, daily, weekly or monthly, in the user's time zone. A 60-second timer claims a due job with one conditional UPDATE that also sets the next run, so two servers never take the same job and a crash can't make it loop.

2The user's rights, not admin

The job runs as its responsible person. Each run mints a fresh 30-minute token from that person's "AI Insider" API key. Switch the key off in Settings and the agent stops.

3One job, two phases

The manager writes one instruction
A marker splits it into search and CRM parts
OpenAI Responses API, provider set per region
MCP attached as a remote tool, with an allow-list

4Guards and a paper trail

No web search in phase 1: answer rejected
No tool call in phase 2: run marked ERROR
Result counted in the DB, not taken from the model
One log row and an email per run
In chat, Insider asks before mail, money or bulk edits

What testing changed

Invented IDsthe model reported companies it never created, until an "order of work" rule and the no-tool-call guard

Companies from memoryanswers came without a search, so phase 1 now fails unless the web search really ran

Same finds every daya search memory of the last 150 companies keeps each run new

Rate limits13k-token requests hit the per-minute cap; retries now wait as long as the API asks