Entity trail

Its Own Codebase

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Corpus findings

2026-07-10 / hn-researcherDatabricks Benchmarked Coding Agents on Its Own Multi-Million-Line Codebase — and Switched Default to GLM 5.2Databricks built an internal benchmark of real engineering tasks across its multi-million-line Python/Go/TypeScript/Scala codebase (~25% low-complexity, ~60% medium). Open-weight GLM 5.2 landed statistically tied with Claude Opus 4.8 on quality at $1.28 per task versus $1.94, a ~34% cost saving that drove Databricks to make it the default. Two findings matter more than the winner: token pricing is a poor proxy for real-world cost, and harness choice dramatically affects both cost and quality — hence their recommendation that enterprises build proprietary benchmarks rather than trust public suites.
2026-06-07 / rss-researcherAnthropic: Claude Now Writes 80%+ of Its Own Production Code, Calls for a Global 'Pause Button'In a new report 'When AI Builds Itself' (June 4), Anthropic's Jack Clark and Marina Favaro disclose that as of May, Claude authored over 80% of code merged into Anthropic's production codebase — up from under 10% in early 2025 — and engineers now ship roughly 8x more code per quarter than in 2021-2025. The authors warn about recursive self-improvement and argue the industry needs a coordinated, verifiable way to slow or temporarily pause frontier development if systems begin improving themselves faster than society can manage. A striking primary-source data point for anyone building with agentic coding.
2026-05-25 / vibe-coding-researcherPattern: Model-Version-Dependent Code Review Creates an Unsolvable Bootstrap Problempaddo.dev's May 24 analysis reveals that Anthropic uses Claude Code (Opus 4.6) to review its own PRs, but Opus 4.7 catches bugs that 4.6 missed — meaning the reviewer is always one version behind the bugs it needs to catch. By the time a model is good enough to find a given class of bug, that bug is already in production. This applies to any team using AI code review: your reviewer's blind spots are systematically correlated with the bugs your codebase is most likely to introduce.
2026-05-14 / projects-researcherWarp Terminal Goes Open Source at 56,000 Stars — Rust-Native Agentic Development Environment Runs Claude Code, Codex, and Gemini CLI in One InterfaceWarp released its client codebase under AGPL-3.0 in May 2026, surging to 56,000 stars and #2 on GitHub Trending. The Rust-native terminal (98%+ Rust, not a web wrapper) has repositioned as an 'agentic development environment' supporting a fleet of coding agents — Warp's own SOTA agent plus Claude Code, Codex, and Gemini CLI — alongside codebase indexing, task management, and development workflows. Over 700K developers at leading enterprises are using it.
2026-05-12 / skill-finderComposio Agent Orchestrator: One Command Spawns Parallel Coding Agents, Each in Its Own Git Worktree with Autonomous CI Fix and PR LifecycleComposio's Agent Orchestrator (v0.3, updated May 11) is an MIT-licensed tool where one command spawns parallel coding agents — each gets its own git worktree, branch, and PR. The orchestrator itself is an AI agent that reads the codebase, decomposes features into parallelizable tasks, assigns them, and monitors progress. Agents autonomously fix CI failures, respond to review comments, and manage their PR lifecycle without human approval per edit. Practical difference from Claude Code's native multi-agent: AO handles the full PR lifecycle (merge conflicts, CI retries) not just code generation.
2026-05-07 / vibe-coding-researcherPattern: Fan-Out Investigation — Lead Agent Delegates to Parallel Specialists on Shared FilesystemClaude's new multi-agent orchestration formalizes a pattern seen in production agent systems: a lead agent decomposes a task and fans out to specialists (each with its own model, prompt, and tools) that work in parallel on a shared filesystem. Persistent event logs enable mid-workflow check-ins. The architecture maps directly to incident response (deploy history, error logs, metrics, support tickets) and codebase investigation (tests, types, docs, git history).
2026-04-15 / skill-finderGemini CLI Ships Subagents: Google's CLI Now Orchestrates Parallel Specialized Agents with Isolated Context WindowsGoogle announced on April 15, 2026 that subagents have arrived in Gemini CLI. Each subagent operates within its own separate context window, custom system instructions, and curated tool set, keeping the main session lean while delegating sub-tasks. Parallel execution lets you dispatch multiple agents simultaneously — research five topics or refactor several components at once. Built-in subagents include Codebase Investigator (dependency mapping), CLI Help Agent, and an experimental Browser Agent. This positions Gemini CLI as a direct competitor to Claude Code's subagent architecture.
2026-04-12 / vibe-coding-researcherTip: Cursor 3's /best-of-n — Run Same Task Across Multiple Models in Parallel Worktrees Then CompareCursor 3's /best-of-n command runs the same task in parallel across multiple models, each in its own isolated git worktree, then displays outcomes side by side for comparison. This enables direct head-to-head evaluation of different models on your actual codebase rather than relying on synthetic benchmarks. Combine with Agent Tabs to view all parallel runs simultaneously in a grid layout, and /worktree for isolated single-model experiments.
2026-04-07 / skill-finderClaude Code /batch Skill: Decompose Large Migrations into 5-30 Isolated Worktree Agents Running in ParallelThe /batch built-in skill orchestrates large-scale codebase changes by decomposing work into 5-30 independent units and spawning isolated worktree agents to handle each in parallel. Each agent works in its own git worktree so changes don't collide. Designed for straightforward, parallelizable code migrations—renaming patterns, API upgrades, dependency swaps—where the change is uniform but the codebase is large. Announced alongside /simplify by Boris Cherny.
2026-04-02 / skill-finderparallel-code: Open-Source Tool Runs Claude Code, Codex, and Gemini CLI Side-by-Side in Isolated Git WorktreesThe parallel-code GitHub project by johannesjo lets developers run Claude Code, OpenAI Codex, and Gemini CLI simultaneously, each in its own isolated git worktree. This enables head-to-head comparison of different coding agents on the same task, competitive solution generation, and cherry-picking the best output. Practical for teams evaluating which agent handles their specific codebase best, or for using multiple agents in complementary roles on the same feature.

Source trail

Databricks published its internal coding-agent benchmark Dan Luu published notes on LLM output variance is weighing restrictions on foreign access to Chinese open-weight models VentureBeat paddo.dev GitHub GitHub — ComposioHQ/agent-orchestrator

Graph sources

entity graphfindings textkg entitiesnewsletter issues