Claude Opus 4.8 in GitLab: More Precision on Complex Multistep Tasks

Anthropic has released Claude Opus 4.8, a new model built specifically for agents that work autonomously on complex projects. Starting this week, it is available in GitLab Duo Agent Platform, where it can run multistep tasks considerably more accurately and reliably than previous versions. This matters most for teams that have already put agents into their workflows and need more precise output.

Precise execution of complex tasks

Agents often stumble on multistep projects: they lose sight of the original goal, skip critical steps, run them in the wrong order, and make unnecessary rework. Opus 4.8 interprets instructions with much greater precision and runs long sequences of operations without failures, even when the task requires working with several tools or calling different APIs.

The result: teams get a clean final output with minimal human intervention. Every step is carried out as the instructions specify. Far less time goes into checking and fixing the agent's errors, which saves hours on complex workflows, especially when an agent works on a project for several hours.

The model also handles long runs better: if an agent is launched for an extended multistep process, Opus 4.8 holds on to context better and doesn't lose the goal along the way.

Beyond code

Opus 4.8 is good for more than coding and development. It works more reliably with documents, data analytics and knowledge structuring. For teams that use GitLab Duo agents for planning, documentation, analysis and code at the same time, this means better accuracy across the board.

Where the model shows improvements:

• Writing, editing and formatting documents in various formats
• Data analysis, report preparation and creating visualizations
• Structuring and organizing large volumes of information from different sources
• Running multistep workflows across different tools and applications
• Synthesizing and summarizing information from several sources into a single structured output

This widens the range of tasks that can be automated through agents.

Updating instructions on the fly

A new feature: support for updating system instructions in the middle of a session. Previously, if conditions changed during work (files were updated on disk, new context appeared, the token budget changed), you had to reload the session completely and lose the entire cache, which slowed things down.

Opus 4.8 lets you update system instructions without resetting the prompt cache. This speeds up asynchronous workflows: new information arrives mid-run, the system adapts, the cache stays valid, and execution continues without a reload.

It is useful for integrations that deliver data in pieces, or when requirements change during execution, since the system stays in sync with the current state.

What this means

Agents are becoming more dependable for production tasks. Fewer errors and less rework mean less time and money spent on fixes. For DevOps engineers and analysts, this means they can entrust agents with genuinely complex, multistage automation processes without worrying that the model will lose the goal halfway and force them to intervene manually.

And support for updating instructions on the fly means workflows can adapt to changing conditions in real time.