CoreStory has released V3 of its code intelligence platform, which turns large legacy codebases into a knowledge graph of business rules that AI coding agents can use. The San Francisco company announced it on 17 November 2026 and is showing it at Microsoft Ignite. It says agents from OpenAI and Anthropic handled hidden business rules correctly in 40% to 74% of attempts with CoreStory, up from 4% to 36% alone.
Quick Take
- What: CoreStory V3 reads a codebase once and ties each business rule to the exact lines that enforce it.
- Claim: agents fixed or flagged the missing rule in 4% to 36% of attempts alone and 40% to 74% with CoreStory, depending on the model. These are the company’s figures.
- Compatibility: works with any coding agent or harness that supports MCP, including Claude Code, GitHub Copilot, Cursor and Codex.
- Pricing: free Developer plan, Pro at $20 a month, Team at $80 a month, Enterprise on request.
- Availability: from 17 November 2026.
In plain English
Old software often hides rules in its code, such as how a discount or a refund is calculated. A coding agent that does not know these rules can break them. CoreStory scans the code, writes down the rules and where they live, and hands the relevant ones to the agent while it works.
What happened
CoreStory said V3 watches commits to the application and updates its knowledge graph as code changes. It gives the relevant rules to a coding agent while the agent plans and edits, and flags changes that conflict with a known rule. The company says this lets agents draft changes that respect the rules.
Background
Enterprise teams have been giving coding agents larger tasks, but legacy systems with undocumented logic remain a weak spot. Tools that expose project context through MCP, the Model Context Protocol, have become a common way to feed agents that context.
How it works
The platform builds a graph linking each business rule to its code location. Engineers, analysts and agents query the same graph through MCP. When an agent proposes a change, CoreStory checks it against known rules and raises a flag on conflicts.
Why it matters
For enterprises with legacy code, the product targets a concrete failure mode of agents, silently dropping rules nobody wrote down.
For developers, a free tier and an MCP interface make it easy to test with the agent they already use.
What’s still open
- The 4% to 36% and 40% to 74% results come from CoreStory, and the test setup was not detailed in the announcement.
- Even the best reported rate leaves about a quarter of attempts failing.
- How well the graph holds up on very large or poorly structured codebases.
Sources
- CoreStory, “CoreStory V3 Opens the Enterprise’s Most Critical Systems to AI Coding Agents,” Business Wire, 17 November 2026
i10x Perspective
Agents are only as safe as the context they get. Rule-aware context layers like this are a practical way to make coding agents trustworthy on old code, but the numbers need independent testing. Explore the Coding section and our report on shared session history for coding agents.



