Lenny's PodcastAsha Sharma: Why org charts give way to agent work charts
How products become living, learning organisms with the loop at the center; Sharma on post-training, reward models, and work charts replacing org charts.
CHAPTERS
- 0:00 – 1:01
Cold open: Agentic society, near-zero marginal cost, and flatter organizations
Asha frames a future where high-quality output becomes extremely cheap, driving exponential demand for productivity. Agents become the scaling mechanism, and traditional hierarchies compress as “org charts” start to resemble “work charts.”
- •Marginal cost of good output trends toward zero
- •Exponential demand for productivity/output follows
- •Agents become the primary way to scale work
- •Fewer layers needed; hierarchy gives way to task-based structures
- 1:01 – 4:26
Who Asha Sharma is and why her vantage point on AI matters
Lenny introduces Asha’s unusual scope at Microsoft AI Platform and her prior leadership roles. The setup highlights why she sees cross-industry patterns in what works (and fails) when shipping AI at scale.
- •Asha’s role spans infrastructure, foundation models, agent toolchains, responsible AI, and growth
- •Past leadership: Instacart COO; Meta product leader across Messenger/Instagram Direct
- •Unique exposure to many companies building on AI platforms
- •Episode themes previewed: product-as-organism, post-training, agents, UI shifts
- 4:26 – 6:40
From “product as artifact” to “product as organism”
Asha explains the shift from shipping static software to building products that continuously learn and improve through interaction. The new competitive advantage becomes the “metabolism” of a team and system: how quickly it can ingest signals, optimize, and evolve outcomes.
- •Models now tool-call, function-call, and take actions—enabling new product primitives
- •Success shifts from shipping features to running continuous optimization loops
- •Tuning toward outcomes (price, performance, quality) becomes central
- •Products that “think, live, and learn” become a company’s new IP
- 6:40 – 9:12
Post-training becomes the new battleground: optimizing loops over bigger pre-trains
They connect product moats to data and iteration cycles, then zoom into why post-training is accelerating. Asha argues that beyond a certain model scale, it’s often more economical and effective to optimize via fine-tuning/RL and rigorous experimentation than to pre-train from scratch.
- •Economic inflection: training huge models can stop making sense past certain scale (e.g., ~30B parameters referenced)
- •Optimization focus moves to reward design, evals, A/B testing, and iteration
- •Data can be first-party, purchased, or synthetic—what matters is the loop quality
- •Real products require multiple parallel loops, not a single feedback cycle
- 9:12 – 12:06
Patterns of successful AI companies—and the most common failure mode
Asha lays out a pragmatic maturity path: make everyone AI-fluent, improve existing processes, then use AI to drive growth. She contrasts this with “AI for AI’s sake,” where teams launch many projects without measurement, observability, or an adaptable platform layer.
- •Step 1: organization-wide AI fluency (copilots and daily usage)
- •Step 2: apply AI to existing processes (support, fraud cycle time, ops workflows)
- •Step 3: use AI to inflect growth (retention/LTV, new categories, embedded→embodied agents)
- •Failure mode: too many unmeasured projects; lack of evals/observability; no clear blueprint
- •Enterprise guidance: bet on platforms that allow swapping tools as the landscape changes
- 12:06 – 13:50
The renaissance of the full-stack builder (and why org complexity can’t keep up)
Asha argues roles are converging because traditional product shipping has too many handoffs for the current pace of AI change. Organizations that empower polymath, full-stack builders gain throughput and can run tighter learn-and-improve loops.
- •Technology shifts historically create new roles; AI shifts roles again toward polymaths
- •Traditional launches involve many steps, functions, and layers—creating excessive coordination cost
- •Rapid model/tool churn makes slow handoffs untenable
- •Full-stack builders increase velocity and tighten iteration loops
- 13:50 – 14:57
“The loop, not the lane”: a new operating model for teams
They unpack Asha’s principle that functional boundaries matter less than owning an end-to-end optimization loop. Continuous feedback, strong observability, and shared accountability become cultural defaults as products become adaptive organisms.
- •Prioritize end-to-end system loops over narrow functional “lanes”
- •Everyone must understand cost, rewards, system design, and UX implications for humans and agents
- •Observability and evals become a core cultural practice
- •Functions blur as teams optimize living products continuously
- 14:57 – 16:07
Concrete examples: GitHub/Cursor-style iteration and Dragon’s clinical impact
Asha gives examples of loop-based product building, especially in coding and healthcare. She highlights how expert annotation and continuous optimization can dramatically improve acceptance rates and outcomes—often with smaller, cross-functional teams.
- •Coding assistants use ensembles/fine-tuned models across languages/regions
- •Dragon for physicians improved markedly with expert-labeled real interactions (vs. synthetic tuning)
- •Continuous optimization loops can drive large jumps in acceptance/quality metrics
- •Small, loop-optimized teams can outperform larger, layered orgs
- 16:07 – 20:28
UI is shifting: from GUI-first to composable, code-native interfaces
Asha predicts a familiar historical pattern: mature systems move from graphical consoles to code-native, composable interfaces (e.g., SQL, Terraform). In AI, text streams and composability align naturally with LLMs and agents, pushing product makers to optimize for machine-readable workflows—not just pixels.
- •Historical analogies: databases (GUI→SQL), cloud (console→Terraform)
- •Text-first interfaces map cleanly to LLM interaction patterns
- •Future products emphasize composability and collaboration at scale over “canvas” design
- •Chat will matter, but won’t be sufficient; artifacts like docs/email/presentations remain important interfaces
- 20:28 – 26:17
The rise of an “agentic society”: embedded and embodied agents change how work scales
Asha describes an agent-driven future where organizations deploy far more agents than today’s software tools. As agents take on longer-running tasks, task routing, supervision, and quality control become the new management primitives—shifting org design toward throughput and task graphs.
- •Embedded agents: software agents with tools; embodied agents may emerge over time
- •Early examples: assigning PRs to Copilot; agentic SDR workflows
- •Org chart → work chart: hierarchy replaced by task-based routing and accountability
- •Key needs: monitoring, fine-tuning, self-healing, strong evals and governance
- 26:17 – 30:32
How Microsoft uses agents internally: incident response, prototyping, and daily workflows
Asha shares practical internal use cases where agents already improve speed and clarity. The examples emphasize summarization, faster diagnosis, rapid prototyping, and ubiquitous writing/documentation support across teams.
- •Agents summarize live-site incident “bridges” to accelerate understanding and response
- •DevOps workflows are changing via AI assistance
- •Rapid prototyping with tools like Spark; “natural language to prototype” accelerates creativity
- •AI is embedded across writing, documentation, and workflow efficiencies
- 30:32 – 35:33
Planning strategy when everything changes: seasons, OKRs, squads, and slack for the slope
They discuss how to roadmap in a world of constant model and capability shifts. Asha explains Microsoft’s “season” framing (e.g., the current season is agents), combined with looser quarterly OKRs and short squad cycles, plus deliberate slack for unforeseen shifts.
- •Six-month planning can still help alignment, but over-baking is risky in AI
- •Use “seasons” to align on secular changes, customer problems, and what winning means
- •Loose quarterly OKRs + 4–6 week squad goals for agility
- •Leave slack for unplanned work and for riding the technology “slope,” not a snapshot
- 35:33 – 39:31
Platform fundamentals and lessons from giants: the invisible work that makes products win
Asha reflects on how platform basics—reliability, performance, privacy, safety—often matter more than feature polish. She draws lessons from Porch, WhatsApp, and Instacart to show how durable advantage comes from infrastructure and matching/availability engines, not surface-level features.
- •Great platforms share traits with great products; impact often comes from “invisible” work
- •Porch: matching engine (pros, service types, geographies) drove core value creation
- •WhatsApp: phonebook reach, reliability, and privacy beat feature proliferation
- •Instacart: real-time, large-scale catalog/availability infrastructure is the differentiator
- •For AI platforms: data residency, reliability, safety, and retrieval tooling are foundational
- 39:31 – 42:01
Leadership lesson from Satya Nadella: optimism as an operating system
Asha shares her biggest leadership takeaway: optimism is renewable and can generate energy and clarity in turbulent markets. She connects “vibes” to mission-driven commitment—why people choose to dedicate their time and effort to a company’s work.
- •Optimism renews dedication and focus amid rapid change
- •Generating energy and clarity is a leadership force multiplier
- •Mission orientation helps sustain commitment in competitive talent environments
- •Culture elements like growth mindset are reinforced by day-to-day leadership energy
- 42:01 – 44:39
What drives Asha: healthcare impact, human flourishing, and long-horizon problems
Asha describes personal motivation rooted in healthcare advances and societal challenges. She frames AI as a chance to extend human capability and tackle “100-year problems,” emphasizing platform-building as leverage for broad impact.
- •Healthcare outcomes are a deeply personal motivator (family experience)
- •AI can expand capability across the workforce and society
- •She worries about macro trends (e.g., declining fertility) and sees AI as part of solutions
- •Microsoft’s platform mission: enable others to build transformative use cases
- 44:39 – 49:19
Reinforcement learning and the next wave of product technique: optimizing outcomes post-launch
Asha predicts RL and post-training will become a dominant product and engineering discipline as agents act in the world. She clarifies pre-training vs. post-training economics and argues organizations will increasingly adapt models to their goals rather than build from scratch.
- •RL/post-training becomes central as products can act, reason, and learn continuously
- •Investment shifts: post-training spend may match or exceed pre-training over time
- •Adapting models (fine-tuning/RL) provides leverage on price/performance/quality tradeoffs
- •Model diversity: use a system/ensemble of models optimized per task (latency, thinking time, retrieval)
- 49:19 – 57:10
Lightning round: books, media, products, mottos—and TaeKwonDo mindset
Asha shares favorites and personal frameworks, from recommended books to home-renovation tools. She closes with how martial arts shaped her mental discipline, courage, and clarity—traits she sees as essential in product-building.
- •Book recs: 'Thinking in Systems' / systems thinking; 'Tomorrow, and Tomorrow, and Tomorrow'
- •Favorite media: Formula 1; For All Mankind
- •Favorite products: DeWalt power pack; smart-home “middleware” experimentation
- •Life framework shift: from minimizing regret to maximizing option value (health, trust, relationships)
- •TaeKwonDo: mental clarity, courage, and persistence as transferable skills
