- Anthropic says it removed over 80% of Claude Code's system prompt for newer models, with no measurable loss on its coding evaluations.
- Anthropic also advises testing whether a fleet of sub-agents can be merged into one agent that loads context on demand.
- Merge one agent at a time and test each step. In one Anthropic test, a merge cut latency but broke a working process.
Old rules, new models
Every team that builds an AI agent ends up writing rules for it. Each rule is a scar from something that once went wrong. Anthropic now says many of those scars can be removed.
Anthropic engineer Thariq Shihipar wrote that the company deleted more than four-fifths of the system prompt Claude Code uses with newer models, including Claude Opus 5 and Claude Fable 5. A system prompt is the standing set of instructions a model reads before every request. Anthropic reports no measurable loss on its coding evaluations.
This shows that much of an agent's complexity is a record of old model weaknesses. When the model improves, that record can turn into the problem.
Why more rules can make an agent worse
Anthropic says it was overconstraining Claude Code. When its team read transcripts of its own internal use, it found conflicting instructions inside a single request. One said to leave documentation as appropriate. Another said not to add comments.
The model can usually work out what the user wants. But Anthropic says it must think harder when messages overlap and clash.
The old system prompt told Claude to default to writing no comments. Anthropic says that advice was wrong for some prompts, because a user may have their own preferences. It accepted the tradeoff for older models. The new prompt asks Claude to write code that matches the surrounding code's comment density, naming and idiom.
Examples can also backfire. Anthropic says that with its newest models, giving examples of how to use a tool "actually constrains them to a certain exploration space."
One agent in place of a fleet
The Decoder reports that Anthropic's Applied AI team drew a wider lesson from enterprise deployments. A single agent that uses skills, which are packaged workflows loaded when needed, frequently beat both one giant prompt and a set of sub-agents on quality. The Decoder says Anthropic measured the single agent running about twice as fast in one comparison.
One company had a dozen specialised sub-agents inside its customer-facing agent. According to The Decoder, that made a consistent brand experience hard to keep and added latency.
Anthropic calls the pattern "Conway's Law" for agents. Conway's Law is the old observation that systems come to copy the structure of the organisation that builds them. Each team ships its own agent, and the agents pile up without shared standards. Teams then add rules to the system prompt that contradict each other. Anthropic calls the result a "Frankenprompt."
How the single-agent design works
The mechanism is called progressive disclosure. The agent does not carry everything in view at once. It loads information when the task calls for it.
Anthropic moved verification and code review out of the Claude Code system prompt and into skills the agent calls selectively. Some tools are also loaded on demand. The agent searches for their full definitions before using them, so they take up no space until needed.
For CLAUDE.md files, Anthropic advises keeping them light. They should say briefly what the repository is for. Most of the space should go to gotchas, meaning things the agent cannot work out by looking at the files. Long guidance belongs in a tree of files loaded at the right time.
The Decoder adds a limit. Anthropic notes that an agent picks tools less reliably once it has somewhere between 30 and 50 tool definitions to choose from. It recommends named owners for each team's skill library so the libraries do not balloon.
Where sub-agents still earn their place
Anthropic does not say sub-agents are obsolete. The Decoder reports three cases where they still make sense. The first is a single agent whose performance measurably drops as its context window fills. The second is work that has to run in parallel, each task in its own context window. The third is an agent that has trouble choosing the right tool.
Anthropic also warns against merging blindly. In one support-agent test, absorbing a sub-agent made responses faster, yet it broke a process that had been working. The main agent started replying itself rather than handing the question to the right specialist.
Two caveats apply. These findings come from Anthropic's own evaluations and deployments, and the sources do not publish the underlying data. Anthropic also sells the models it says make simpler designs possible. Treat the claim as a hypothesis to test in your own setting.
Questions for your team
Ask how many agents your organisation runs and who owns each one's system prompt. Ask whether any instructions contradict each other across the prompt, skills and user requests.
Ask which rules were added for a weaker model and whether anyone has tested removing them. Ask for a before-and-after evaluation on your own tasks, not a vendor's.
If you merge agents, do it one at a time and check the result after each step. The Decoder reports that Anthropic has put a migration kit on GitHub to guide the work. The support-agent test is the reason to verify every merge.
An agent's rules should be reviewed as often as its code. A rule that once prevented a failure can later become the cause of one.
Produced by the WebPulse Newsroom with AI assistance from the original reporting credited below, and checked against that source by our editorial review. How we use AI.
Original reporting: Anthropic.
