TechCrunch’s framing for its Disrupt 2026 session with Jas Khaira, global head of Blackstone N1, is almost disarmingly plain: he will talk about what Blackstone looks for when backing category-defining companies, how founders should think about capital as they scale, and what separates lasting businesses from early traction. That last clause is the one I keep rereading. Not “what separates winners from losers.” What separates a business from a spike.
From where I sit — writing about agent internals rather than cap tables — that distinction is not primarily a financial question. It is an architecture question wearing a finance costume.
Early traction is cheap in agent-land
Here is the uncomfortable property of agentic products in 2026: the demo-to-revenue path has gotten shorter than the demo-to-reliability path. A well-prompted orchestration layer sitting on a frontier model can produce a genuinely impressive first-run experience in a weekend. Retrieval, a few tools, a planner loop, some retries. Users sign up. Logos go on the website. Traction, as measured by the usual dashboard, looks real.
What that dashboard does not measure is whether the system degrades gracefully. And agent systems degrade in ways that traditional software does not:
- Compounding error across steps. A chain with twelve tool calls and 97% per-step reliability lands near 70% end to end. Users do not experience that as “mostly fine.” They experience it as untrustworthy.
- Silent distribution shift. The prompt that worked on last quarter’s user queries quietly stops working as the user base widens. Nothing throws an exception.
- Upstream model churn. A provider deprecates a checkpoint or retunes behavior, and your carefully calibrated planner changes personality overnight.
- Context rot at scale. Long-horizon agents accumulate state, and the failure mode is not a crash but a slow drift into incoherence.
None of these show up in a seed-stage metrics review. All of them show up in year three.
Capital as an architectural input
The second thread in Khaira’s session description — how founders should think about capital as they scale — reads differently when you treat it as an engineering constraint rather than a funding strategy.
Money in an agent company does not buy headcount first. It buys three things that compound: evaluation infrastructure, proprietary interaction data, and inference economics. Each one is slow, unglamorous, and nearly impossible to retrofit cheaply.
Evaluation is the moat nobody screenshots
I would argue that the single most reliable signal of a durable agent company is the maturity of its eval use. Not benchmark scores — internal, task-specific, adversarially maintained evals that run on every change and encode what “correct” means in that domain. Building one requires paying humans to label disagreements, maintaining golden datasets, and resisting the temptation to optimize the metric instead of the product. It is pure cost for several quarters and then becomes the thing that lets you swap models, refactor planners, and ship weekly without fear.
Teams without it are not slower. They are structurally unable to improve, because they cannot tell improvement from regression.
Inference margins are a design decision
Gross margin in an agent product is set largely at architecture time: how many model calls per task, which tier of model handles which subtask, how aggressively you cache and distill, whether routing is learned or hardcoded. A company that treats token spend as a line item to renegotiate later is making a bet that costs will fall faster than usage grows. A company that treats it as a design problem builds routing, small-model fallbacks, and result caching into the system early, then watches margins expand with volume instead of contracting.
Capital can absorb bad unit economics for a while. It cannot fix an architecture whose cost per task scales linearly with user ambition.
Why the theme fits the moment
Disrupt 2026 runs October 13–15 at Moscone West, and the stated theme is building enduring companies in the AI era. “Enduring” is a notable word choice for a conference that has historically rewarded velocity. It suggests the center of gravity has moved from “can you build this” to “can you keep it working for a decade while the substrate underneath you changes every six months.”
That is the real technical problem of the agent era, and it is one Blackstone-scale capital is unusually positioned to notice. Growth investors see companies after the demo glow fades, when the question is whether revenue retains, whether costs bend, whether the team can ship into complexity. The failure modes they underwrite are the ones engineers create.
I am curious whether Khaira talks about any of this in systems terms or keeps it at the level of markets and management. Either way, founders in the room should translate the questions into their own stack. When someone asks how you distinguish lasting businesses from early traction, the honest answer lives in your eval suite, your cost-per-task curve, and your plan for the day your model provider changes its mind.
🕒 Published: