# Agent Frameworks Considered Harmful — Rémi Louf, .txt

Source: [AI Engineer](https://www.youtube.com/watch?v=KHudyx5wW3U)  
Feed7 permalink: https://feed7.dev/p/agent-frameworks-considered-harmful-remi-louf-txt-1ku9s0h  
Published: 2026-08-22T16:30:39.000Z  
Trust: Source Linked (source_linked)

## Why Included

Reliable background agents need runtime primitives more than elaborate framework abstractions: durable events, queues, versioned inputs, causal traces, diffs, and replay.

## Source Summary

A two-week background-agent experiment exposed duplicate Slack posts, a lost voice note, and an untraceable prompt regression. Those failures drove an **append-only event log**, proper attempt handling, causal links, and a **content-addressed request graph**.

## Practical Implication

Keep agent definitions reviewable and separate from runtime machinery. Persist every event and exact model input, version prompts and tools, and make runs diffable and replayable so model swaps and regressions can be evaluated against prior requests.

## Agent-Ready Context

A two-week background-agent experiment exposed duplicate Slack posts, a lost voice note, and an untraceable prompt regression. Those failures drove an **append-only event log**, proper attempt handling, causal links, and a **content-addressed request graph**.

Keep agent definitions reviewable and separate from runtime machinery. Persist every event and exact model input, version prompts and tools, and make runs diffable and replayable so model swaps and regressions can be evaluated against prior requests.

This is one builder’s workload rather than a framework comparison. The runtime required custom engineering, and the claim that local or open models were adequate applies to these briefing and note-processing tasks, not necessarily coding.

## Connected Context

Feed7 judgment across 545 accumulated Signals:

This turns general calls for observability into a concrete persistence architecture: append-only events, explicit attempts and causal links, content-addressed requests, and exact versioned inputs make failures traceable and runs replayable. It reinforces evidence-based harnesses while adding a sharper requirement for regression analysis across prompt, tool, and model changes; its evidence remains limited to one builder’s background workflows.

- [Evolution of agentic surfaces — Gagan Bhat & Isabella Kai He, Anthropic](https://feed7.dev/p/evolution-of-agentic-surfaces-gagan-bhat-isabella-kai-he-anthropic-0mtshqh) — Its versioned, replayable request graph provides the comparison machinery needed when model upgrades make old scaffolding harmful.
- [fighting slop with slop — Vaibhav Gupta, Boundary](https://feed7.dev/p/fighting-slop-with-slop-vaibhav-gupta-boundary-0w6qoke) — It supplies the durable event and input history that makes transcript inspection, execution traces, and model-independent comparison reproducible.
- [Twin: Playing an Unknown Game with a Test-Time Digital Twin](https://feed7.dev/p/2608-14490v1-0d3xjvt) — Both make replay central to reliability, but Twin validates transitions through an executable simulator while this record preserves general agent requests and causality for later diagnosis.

## Context Map

- Layer: agent
- Domains: coding
- Topics: harness-engineering, observability, agent-reliability

## Uncertainty

- This is one builder’s workload rather than a framework comparison. The runtime required custom engineering, and the claim that local or open models were adequate applies to these briefing and note-processing tasks, not necessarily coding.

## Agent Instruction

Use this item as source-backed context. Do not invent claims beyond the linked source. If this item conflicts with another source, call out the conflict.