Claude Sonnet 4.5: why the agent harness matters as much as the model

Anthropic announced Claude Sonnet 4.5 on 29 September 2025. The launch combined a coding-focused model with changes to Claude Code, new context-management capabilities in the API and the Claude Agent SDK. That combination matters more than treating the release as a standalone intelligence upgrade. A model proposes actions; the surrounding software decides which tools it can use, what information remains visible and how a person can intervene. This is a historical analysis published on 25 October 2026, not a claim that Sonnet 4.5 is a newly released or currently recommended model. 1

The announcement also introduced code execution and file creation in Claude conversations. Those application features should not be confused with the model itself: different products can wrap the same model in materially different controls.

Anthropic described sustained performance on lengthy development tasks and presented coding benchmarks as evidence of progress. These are supplier-reported evaluations, not a promise that an unattended agent can safely maintain a production application. Over a longer run, a mistaken assumption can influence more files before anyone notices. A successful demonstration therefore raises a second question: how cheaply can the team discover that the agent is solving the wrong problem?

A useful pilot would ask for a bounded refactor with explicit invariants: preserve the public interface, maintain error handling and change only the named component. Require intermediate evidence rather than accepting a polished final explanation.

Alongside the model, Anthropic announced context editing and a memory tool. Context editing removes stale tool material as a conversation approaches its limits; the memory tool lets an agent maintain information outside the immediate context. Neither mechanism guarantees that what remains is correct. A durable note can preserve a mistaken assumption just as effectively as a useful fact.

Teams should separate verified repository conventions from provisional discoveries. Store the origin of important instructions and review persistent memory when requirements change. The broader distinction between access to information and reliable knowledge is explored in company knowledge for AI.

The Claude Agent SDK exposed core tools, context management and permission frameworks used by Claude Code. Anthropic also described support for subagents and hooks. This made it possible to build more specialised workflows without inventing every orchestration component. It did not make a custom compliance or development agent automatically reliable.

A builder still has to define allowed repositories, tool credentials, escalation points and termination rules. Subagents can divide work, but they can also create conflicting assumptions. Treat each delegated result as an input requiring validation. Use the separate agent permissions analysis for the cross-product security framework rather than duplicating it here.

The accompanying Claude Code announcement introduced checkpoints, a refreshed terminal interface and a native VS Code extension in beta. Checkpoints make experimentation easier by supporting a return to an earlier conversational or code state. They should not be interpreted as a transaction spanning every connected system. An external message, remote update or other side effect may need a separate recovery process.

Before granting write access, list what can be restored locally and what cannot. Run the first experiments in an isolated branch with no production credentials. The acceptance test is recoverability as well as task completion; AI advisory can help translate that distinction into an operating decision.

A practical evaluation should record accepted changes, reviewer time, failed attempts and defects discovered after approval. Compare similar tasks under the same test environment. Long uninterrupted runtime is not a business outcome if the result creates an expensive review queue. Record the model snapshot and configuration so that a later comparison does not silently test a different system.

The lesson from this release is architectural: better models create value when the harness makes their work observable and reversible. For a shared management vocabulary, use an AI keynote, then make the investment decision on repository-specific evidence rather than a launch superlative.