We've noticed that nearly every Agent product in 2025 faces the same awkward problem: the longer the conversation, the more AI drifts off course. After three rounds of revisions to a PPT, the Agent can no longer tell which version is current or which piece of feedback has been voided. The tech community is now reflecting—the problem isn't that the model has poor memory, but that using "conversation" as the primary context is itself flawed.
What This Is
The author calls the current Agent working method "conversation-centric": the user states requirements, the AI reviews historical messages, then makes changes. But conversation is a process log—requirements, explanations, attempts, failures, and feedback all pile up together, and information density thins out the longer the chat goes.
The opposite is "artifact-centric": each round, the Agent first reads the current file (PPT, code, spreadsheet), getting the "actual current state," then combines the user's new request for that round to modify it, and verifies through the runtime environment after editing.
Why are Coding Agents (AI assistants that write code, such as Cursor and Devin) generally more reliable? Beyond the model itself being strong, the more critical reason is that code itself is an "AI-friendly" artifact—text-based structure, searchable, locally editable, with Git version control (records of each modification) and tests for verification. PPTs, Word documents, and 3D models are far more troublesome: getting AI to understand page structure, visual relationships, and spatial state, then having the runtime environment render them out—there is currently no mature "read—modify—display—verify" mechanism.
Industry View
This line of thinking aligns with the Agent field's consensus over the past year: Context Engineering (i.e., how to prepare "working memory" for AI) matters more than prompt engineering, and "what to use as working memory" is the core question. Anthropic and LangChain are both exploring ways to let Agents read external state directly and reduce dependence on long conversations.
But dissenting views are worth flagging. First, the ideal solution described in the article requires every type of artifact to have its own dedicated parsing and verification layer—a massive engineering effort, more suited to companies like Microsoft, Google, and Adobe that "hold the file formats." Independent Agent startups may actually fall behind. Second, chat-centric Agents are popular precisely because most users think by talking it through; forcing artifact-centric design would sacrifice usability. Third, treating artifacts as the sole source of truth would mask the judgments and trade-offs embedded in conversation—this tacit knowledge (colloquially known among engineers as "tribal knowledge") is exactly what Agents find hardest to replicate.
Impact on Regular People
For enterprise IT: If Agents truly go "artifact-centric" in the future, enterprises need to establish versioning and machine-readable access permissions for core files like PPTs, contracts, and engineering drawings. Otherwise, AI won't get the "current actual state" and will fall back into the old chaos of conversation-driven confusion.
For individual professionals: In the short term, don't expect that chatting more rounds with AI will improve your PPT revisions. A more effective approach is to give only one round of incremental requirements at a time, and hand the current file directly to the AI as a starting point.
For the consumer market: Watch the next 12 months to see whether Office, Google Workspace, Feishu, or DingTalk will launch "artifact-aware" AI—able to directly read and modify your documents, rather than just chatting alongside you.