---
title: Five hallway claims about personal agents, checked
description: A builder on the AGNTCon + MCPCon NA livestream said today's agents are immature, the harness beats the model and memory is really about forgetting. He's mostly right. Two of his five claims go too far.
date: 2026-10-25
tags: agents, memory, harness, reliability
kind: short
---

On the AGNTCon + MCPCon NA livestream, a builder named Sid, just retired from big tech, gave five hot takes on personal agents in about four minutes. I checked each one against what I run at home. Three hold up as he said them. Two point the right way but go too far: the harness doesn't beat the model, and failures don't compound "exponentially".

It's a hallway chat, not a talk, and he never names a product. The only pitch is a soft one: this is why my assistant will be different. [Watch it from 2:09:50](https://youtu.be/iNV5-YkPHW4?t=7790).

## One-off tasks work, daily ones drift

His point: booking a restaurant is one task you check once. Answering your email every day isn't. He's seen two emails that contradict each other get two separate replies, with nothing joining them up.

Right, and I've lived it. My daily email digest agent listed about 77 mails a day, and around 5 of them mattered. I deleted it this week. One good run is easy. The same quality every morning is the hard part.

## The harness matters, and so does the model

He said the harness is more powerful than the model, and a model on its own is pretty dumb. The second half is true: the same model swings a lot depending on its tools, its loop and its checks. That's the whole idea behind [forking your agent's harness](/notes/fork-your-agents-harness.html).

The first half goes too far. No harness saves a weak model on a hard reasoning task. The fair version: most people underinvest in the harness.

## Memory is about forgetting

Correct, and the host called it the hardest part. A bigger context window doesn't decide what to keep, what went stale, or which of two facts wins when they clash. My coding agent keeps its memory as small files it updates and deletes, and treats every recalled one as possibly out of date.

## One reliable agent before five talking ones

He wants one agent working consistently before any agent-to-agent setup. The math is the old one the host brought up: if each step is right 98% of the time, ten steps in a row land about 82% of the time. That's multiplication, not an explosion, and checks and retries win a lot of it back. Still the right order to build in.

## Context bloat is real, and has fixes

Teaching an agent every tool in its prompt costs tokens on every call. He put it at 5,000 to 10,000. True, but he skipped the fixes. Prompt caching makes a repeated prompt much cheaper, and tools can load only when the agent asks for them. The session I used to check his claims loaded most of its tools that way.

## Try the boring test

None of this is new, and that's fine: it's a fair snapshot of personal agents in late 2026. If you're building one, start with his first point. Give it the same job every day for a week, then read the outputs side by side. Tell me what drifts.