# How I run my companies with 26 AI agents

> What the agent organisation behind Northlane actually does each day, what works, and what has broken.

13 Oct 2026, by Ali Asif.

I have 26 agent definitions in a folder on my Mac. Thirteen work on Northlane's bids, clients and compliance, eight make up a small accounting venture desk, and the rest do odd jobs. Before anyone pictures a swarm: they're prompt-defined roles with persistent Markdown memory, run by Claude Code. There's no trained model, no agent framework in the middle. Each one is a written role with a memory file.

## What they do

Each client account gets its own persona that holds the tender pack, the open questions and the cost model for that bid. An estimator sits above them and owns the price memory, so no persona invents a rate. A catalogue agent answers "who can supply this and what did it cost last time". A registrar answers company-record questions from one master file and cites the document. A chief-operating-officer agent takes my loose requests and hands them to the right one.

A person signs off every price and every outbound document. I'm the person. That one rule is why I trust the rest.

## What works

**Written roles beat clever prompts.** The agents that behave are the ones whose job, inputs, outputs and stop conditions I wrote down like a job description.

**Cheap models read, the expensive one judges.** Haiku classifies a thread, Sonnet extracts and drafts, Opus is saved for risk challenge and verification. It's faster and the bill stays sane.

**Voting fixes reading.** On a 99-page scanned Arabic tender, one reading pass agreed on only 30% of part numbers. Rendering at 600 dpi and letting independent readers vote got me 4,362 usable line items.

**Memory files get better with use.** Each persona appends what it learned to its own master file, so the next bid starts further along.

## What breaks

**Scripts drift.** A tracker script once pulled an unrelated supplier's thread into the wrong row because it matched on subject lines. I anchored it to stored thread IDs and checked all 66 rows.

**Agents grade themselves generously.** On one review task a model scored its own draft about 11 points above a blind panel. It was one task, not a study. But now nothing I care about ships on its own self-review.

**Discovery tools return junk.** My contact-finding tool once treated product-page titles for electrical parts as company names and created bogus supplier rows. I now check what it writes before anything gets used.

**Old numbers outlive their truth.** 77 of the 236 rows in my pricebook are past their validity date. They're flagged, but the temptation to quote "all rates current" is real, so I don't.

## What I'd tell you to copy

Start with one agent and one real workflow. Give it a human approval step before it touches anything outside your building. Log every hand-off. Then measure it against something blind, not against your own impression.

About 90% of my own operating workload runs this way now. That's an operating-model figure I stand behind, not an audited one. The day I can back it with logs of hours per bid, I'll say so here.

Canonical: https://heyaliasif.com/writing/2026-10-13-how-i-run-companies-with-26-ai-agents