Pinecone's Edo Liberty argues agents need a persistent knowledge layer for tribal knowledge

Social·Community

If we want them to do Knowledge Work, design them as Knowledge Agents — Benjamin Clavié, Mixedbread

In a talk, Mixedbread's Benjamin Clavié argues the industry generalized agent design from coding, the domain least resembling other knowledge work: code carries durable, greppable cues while general knowledge work has implicit meaning and domain-dependent strings. He traces historical tool and organization loops humans used to solve this and cites benchmark numbers showing well-tuned retrieval and searcher-agent layers closing much of the gap to human performance. The practical takeaway for practitioners is to design agents as knowledge agents with better retrieval tooling rather than assuming coding-agent patterns transfer.

In a talk, Mixedbread's Benjamin Clavié argues the industry generalized agent design from coding, the domain least resembling other knowledge work: code carries durable, greppable cues while general knowledge work has implicit meaning and domain-dependent strings. He traces historical tool and organization loops humans used to solve this and cites benchmark numbers showing well-tuned retrieval and searcher-agent layers closing much of the gap to human performance. The practical takeaway for practitioners is to design agents as knowledge agents with better retrieval tooling rather than assuming coding-agent patterns transfer.

Knowl is an agent memory system that not only stores facts but also updates or removes them when they become stale, preventing AI agents from giving conflicting answers. It solves the problem of long-term context management for AI agents, ensuring they always use the latest information. Aimed at developers using AI agents in production.

Social·Community

Peter Werry (Unblocked) argues code agents mistake search results for codebase understanding

In a YouTube talk, Unblocked's Peter Werry describes how coding agents fail on real codebases: they search, find one plausible answer, and stop — a failure mode he likens to radiologists' 'satisfaction of search.' He argues access to information is not understanding and frames context engines as the fix, replacing the human role of carrying tribal knowledge.

In a YouTube talk, Unblocked's Peter Werry describes how coding agents fail on real codebases: they search, find one plausible answer, and stop — a failure mode he likens to radiologists' 'satisfaction of search.' He argues access to information is not understanding and frames context engines as the fix, replacing the human role of carrying tribal knowledge.

Social·Community

PromptQL's Tanmai Gopal: scoped-wiki company brains with human-approved writes for Fortune 100 banks

In a YouTube talk, PromptQL's Tanmai Gopal describes plotting daily edits to his company's internal knowledge system (a "company brain"). He expected the typical burst-then-decay usage curve, but edits kept climbing. He reads sustained growth as a sign of trust: as users teach the system new skills like querying, interpreting, and acting on data, each skill adds its own stream of corrections.

In a YouTube talk, PromptQL's Tanmai Gopal describes plotting daily edits to his company's internal knowledge system (a "company brain"). He expected the typical burst-then-decay usage curve, but edits kept climbing. He reads sustained growth as a sign of trust: as users teach the system new skills like querying, interpreting, and acting on data, each skill adds its own stream of corrections.

An arXiv position paper argues enterprise AI agents stall in production because they reason over data structured for humans, not for language models. It proposes rebuilding a company's 'cognitive substrate' around agent-matched representations, with schema translation isolated at the action boundary. The paper offers a four-layer framework (Data, Knowledge, Intelligence, Governance) with a Sync Agent and per-skill trust gradient, and outlines a research agenda for evaluating substrates directly.