deep read: @ashpreetbedi
2026-02-19
founder of agno — multi-agent runtime. prev airbnb, facebook.
17 articles, jan–feb 2026. read chronologically, this is one long pitch disguised as a blog. but inside the pitch are 3–4 ideas that genuinely matter for anyone building agent systems. the trick is separating the signal from the agno marketing.
the good
the 3-layer knowledge architecture is the single best idea in this entire corpus — across all five authors. he breaks knowledge into three storage strategies: static context always in the prompt (risk limits that must never be missed by search), vector search for large corpora (company profiles where semantic retrieval shines), and file browsing for structured documents read whole (investment memos that lose meaning when chunked). the insight isn't just "use different storage." it's that where knowledge lives determines how reliably agents use it. a risk limit in a vector store is a liability. a risk limit in the system prompt is a guarantee.
"memory is learning" is a genuine conceptual reframe. memory is a noun (static database of facts), learning is a verb (evolves, compounds, gets sharper). the distinction between extraction (pulling facts from conversation) and integration (teaching the agent to use those facts naturally) is the part most memory system builders skip. claude's memory feels magical not because it extracts well, but because it integrates invisibly. ashpreet names this gap clearly.
the investment committee article is a masterclass in multi-agent design. five orchestration patterns (route, broadcast, coordinate, task, workflow) applied to the same 7 agents with honest trade-offs. the broadcast pattern — independent analysis with no anchoring bias, then chair synthesis — is the cleanest articulation of why you'd want agents that can't see each other's work.
the claude code workflow in "the shift" is immediately actionable. symlinked specs directory, layered CLAUDE.md files (repo-level for navigation, feature-level for constraints), ADR records for decision traces. the 10-minute PR rule enforced in CLAUDE.md is the kind of concrete constraint that actually changes output quality.
the bad
the agno marketing gets heavy. every diagnosis leads to the same prescription: use agno. the runtime thesis is well-articulated, but the conclusion that only agno can solve it is the part you should ignore. the "pitch vs reality" article warns against the eval-industrial complex while pitching its own evaluation approach. the observability series reads like product docs wearing a blog's clothing.
verdict
read the investment committee piece and the memory-is-learning thesis. skip the product launches. the 3-layer knowledge architecture alone is worth your time.