---
title: 'In the News: August 13, 2026'
description: 'Datadog deleted a year of AI context files and its evals improved, reframing agent context work as pruning as much as writing.'
canonical_url: 'https://darkfactory.dev/news/2026-08-13-morning'
markdown_url: 'https://darkfactory.dev/news/2026-08-13-morning.md'
collection: news
date_published: '2026-08-13T08:35:00-04:00'
date_modified: '2026-08-13T08:35:00-04:00'
---

# In the News: August 13, 2026


Datadog deleted a year of hand-written AI context files from one of its repos, expecting worse agent performance. Its evals improved instead. Simon Boudrias, who runs Language Foundations at Datadog, described the finding on the AI Native Dev podcast this week. It is this archive's clearest data point in the argument that steering documents rot as models improve. Two other pieces examine what a coding agent's harness must verify and where quality lives once nobody has time to read the diff.

## 1. Datadog deleted a year of AI context files. Its evals got better.

**[Datadog Deleted All Its AI Context. It Worked.](https://tessl.io/podcast/117)** · Simon Boudrias, Language Foundations lead, Datadog · The AI Native Dev podcast (Tessl), Aug 4, 2026

Datadog scaled Cursor and Claude Code to its roughly 4,000-engineer org in under a year, moving from a 100-to-200-person pilot to more than a thousand daily users within a month of launch. Rather than write more rules, Boudrias's team invested early in evals, anchoring correctness to real production incidents: they replayed pull requests that had caused past outages and measured whether an AI reviewer would have caught the problem before it shipped. That habit paid off when the frontend team questioned a steering document written roughly a year earlier for Sonnet 3.5. They deleted it outright and reran their evals. Performance went up, not flat. Boudrias's explanation: the document had become what the episode calls context rot, instructions a weaker model needed that now compete with more relevant context for a frontier model's attention. Datadog's own eval data was what let the team defend the deletion; without it, Boudrias said, cutting anything from an AGENTS.md file "would be pretty controversial."

**Why it matters:** The instinct with agent context is almost always to add more. Datadog's account is first-party evidence that the more common failure is context nobody has gotten around to deleting, and that the fix is measurable rather than a matter of taste, provided a team has evals to check the deletion against.

## 2. Anthropic's long-horizon agents put the verifier in its own context window

**[Claude for Long-Horizon Tasks](https://www.youtube.com/watch?v=9QebvrrY3KY)** · Lance Martin, Anthropic · AI Engineer conference talk, filmed late July 2026

Martin's talk lays out the architecture behind Anthropic's newer managed-agents surface for long-running work. The harness running the agent loop is decoupled from the containers executing its work and connected through an append-only session log, so a crashed container does not lose the session and credentials never sit inside the sandbox itself. Verification also runs in a separate context window from the one that did the work: when the same context both produces and grades an answer, Martin described the results as unreliable and prone to fabricated self-assessment. Splitting build and verify into a loop, tested on OpenAI's parameter golf ML-research benchmark, produced markedly stronger results from frontier-capacity models. On memory, Martin described an offline consolidation pass, modeled loosely on how the brain moves short-term memory into long-term storage during sleep, that corrects errors an agent wrote to its own memory during a task. In a repeated test using an agent playing Pokemon, an uncorrected memory error caused the agent to fall through the same in-game trap door in five separate runs; runs that went through the consolidation pass avoided the error.

**Why it matters:** Teams could add session logs decoupled from execution and verification kept out of the working context without waiting on new model capability. The memory-consolidation piece is earlier-stage, but it targets a failure mode most long-running harnesses do not yet check: an agent's own written memory going wrong and staying wrong.

## 3. Where code quality goes when nobody has time to read the diff

**[Agentic Code Quality](https://addyo.substack.com/p/agentic-code-quality)** · Addy Osmani · Elevate (Substack), Aug 8, 2026

Osmani's argument is that code review does not scale to agent-generated volume, so quality has to move into the constraints placed around the agent: tests, mutation testing, complexity and line-length limits, and whatever else gates a change before it reaches production. He lays out a three-lever model for what to do when verification capacity is outstripped by agent output: scale the verification system, throttle how fast agents generate changes, or lower the quality bar, and argues teams should expect to use all three at different times rather than pick one permanently. "Quality is in the constraints that we place around our agents," he writes, closing on the idea that the design question is no longer what a human catches on review, but what the harness is built to catch before a human ever sees it. The essay also circulated as a long-form X post that had drawn over 360,000 views by the morning of the 13th, an unusually wide reach for this kind of piece.

**Why it matters:** This is a framing piece more than a new technique, but it names a decision every team running agents at volume is already making by default, whether deliberately or not: which checks are load-bearing enough to gate a merge, and which are just process.