Anthropic Code w/ Claude Conference: How AI Cut a 10‑Week Project to 4 Days

Model evolution and performance benchmarks

Anthropic has released 18 Claude model versions in the past year, spanning Haiku, Sonnet, Opus and the latest Mythos. Each generation adds tool‑use capability, safety improvements and higher test‑time compute. Opus 4.7, released in the last month, outperforms prior models on internal benchmarks and on external customer tests:

Rakuten reported a three‑fold increase in the number of production engineering tasks completed with Opus 4.7.

Intuit observed that Opus 4.7 can automatically detect and correct its own logical errors during the planning phase, reducing iteration cycles.

Amplify (referred to as “Amp”) migrated all of its “smart patterns” to Opus 4.7, eliminating much of the scaffolding and tooling previously required.

Mythos recently parsed the entire OpenBSD source tree and identified a vulnerability that had persisted for 27 years.

Claude Platform new primitives

Three primitives were added to the Claude Platform to enable more autonomous and goal‑driven workflows:

Managed Agents – an enterprise‑grade managed‑agent runtime that binds intelligent agents to production infrastructure. Teams report up to a ten‑fold acceleration from prototype to production when using Managed Agents.

Outcomes – a declarative goal definition format (Markdown) that lets developers specify success criteria. The platform spawns a “grader” agent that continuously evaluates progress against the defined outcomes and iterates until the criteria are met.

Dreaming – a self‑learning memory update mechanism. A dedicated dreaming agent reviews past sessions, extracts lessons, and writes them into a memory store that is automatically consulted by future runs.

The platform also introduced an Advisor strategy . Developers add a low‑cost model (e.g., Haiku or Sonnet) to the tools array of the Messages API; when the low‑cost model encounters a roadblock it forwards the request to a higher‑cost model (e.g., Opus) for guidance. Eve Legal demonstrated that this pattern achieved comparable quality at one‑fifth of the original cost.

Claude Code desktop environment

Claude Code is now available as a desktop application that integrates the Claude Agent SDK and adds several developer‑facing primitives:

Routines – event‑driven automation that can be triggered by webhooks, scheduled timers or API calls. Routines can launch Claude Code sessions in the background, enabling fully asynchronous workflows.

Claude Security – a nightly scanner that inspects an entire codebase for vulnerabilities and automatically invokes Claude Code to generate patches.

Auto‑fix – a real‑time listener for CI failures, code‑review comments and merge conflicts; when a problem is detected the system proposes and applies a corrective patch without developer intervention.

Code‑review bots and remote‑control extensions for iOS and Android allow developers to start or monitor Claude Code tasks from any device.

Rate limits for Claude Code and Claude Platform tiers have been doubled for Pro, Max, Team and Enterprise plans, and API limits for Claude Opus have been substantially increased.

Compute capacity has been expanded through a partnership with SpaceX’s Colossus 1 data center, with the additional resources earmarked for independent developers and small teams.

Customer case studies

Stripe’s infrastructure team reduced a 50 k‑line Scala‑to‑Java migration that was estimated to take 10 weeks down to 4 days using Claude.

Binti’s foster‑care platform shortened the eligibility‑approval workflow by 20 days, enabling children to be placed in homes earlier.

Shopify reported a 200 % increase in pull‑request output after rolling Claude Code across engineering, product, design and data teams, with no measurable drop in code quality.

Mercado Libre processed over 500 k PR reviews and refreshed more than 9 000 applications using Claude Code; the team set a target of 90 % automation of PR cycles for the current quarter.

Eve Legal achieved comparable legal‑document quality while spending only 20 % of the original cost by employing the Advisor strategy.

Demonstration of multi‑agent orchestration, outcomes and dreaming

A live demo built a fictional startup “Lumara” that plans autonomous lunar‑drone landings. The configuration used three agents:

“Commander” – coordinates the mission and aggregates results.

“Scout” – identifies high‑purity mineral sites for landing.

“Navigator” – pilots the drone to the selected site.

The commander was set as the parent session, while the scout and navigator each ran in independent threads with separate context windows. An Outcomes markdown file defined hard success criteria (soft landing, flat terrain, sufficient return‑fuel). During execution a “grader” agent continuously evaluated the outcomes. After the first run, two of six candidate sites failed to meet the criteria. The team then activated Dreaming via the Claude console: the dreaming agent ingested the failed simulations, generated a “descent playbook” with heuristic improvements, and wrote the insights to the memory store. A second run using the updated memory succeeded on all six sites, demonstrating rapid hill‑climbing optimization with a single button press.

Shift to asynchronous, background development

The conference emphasized moving from synchronous, manual prompt‑writing to asynchronous, background execution. Developers can now configure Routines to listen for repository events (e.g., new Jira tickets), automatically invoke Claude Code, and let the system generate, test, and merge code while the developer works elsewhere. Auto‑mode, Managed Agents and the new primitives together enable a development flow where the model continuously runs, validates and improves its own output without constant human prompting.