Why
An agent harness — sandboxing, memory, scheduling, tool execution — is expensive to build and easy to mistake for the product. The panel's founders all started by asking what layer their users actually value: Saumil's users value a briefing that is right about who they're meeting, not the infrastructure that produced it; Mihir's users value a forecast rolled up across hundreds of accounts, not the container it ran in. Getting this wrong in either direction is costly: build the harness when you don't need to and you spend your runway on plumbing instead of product; buy it when you actually need extreme control (sub-second voice latency, cost-optimized batch inference at scale) and you inherit someone else's ceiling. The panel's answer is concrete enough to apply directly: ask "is this infrastructure our core competency?" (12:27) before every build decision, not after.
How it works
Three products, one production pattern
Saumil's meeting assistant does more than live transcription: it briefs the user before a meeting and automates tasks after it (00:11, 00:17). Twenty-four hours ahead, it cross-references the other side's LinkedIn, prior conversations, and email to write the brief (02:27, 05:29). Mihir's Watchtower runs a persistent agent per sales account (00:46, 00:53), and a second agent sits above them, cross-analyzing hundreds of accounts for roll-ups, revenue forecasts, and priority targets — built and deployed on Managed Agents in two weeks (01:14, 01:28). Todd's product combines a customer's source code with their user telemetry: it reviews UX and instrumentation when a PR is submitted, and when funnel conversion drops, it analyzes the code and proposes a fix PR (01:32, 01:53, 23:01, 24:04).
An independent verifier that prefers silence to a wrong answer
All three treat wrong output as worse than no output. Saumil's clearest failure mode is briefing the user on the wrong person — a name collision, the wrong "Tom" — which is far more damaging than skipping the brief entirely (02:18, 02:41). Their fix is a verifier agent that runs in a clean context, fully separated from the agent that did the work (03:10, 06:12), scoring the output against a rubric written from the user's experience, not the model's confidence. Below the rubric's threshold, the system doesn't show a degraded answer — it shows nothing, a fail-silent / graceful-degradation choice made deliberately at the product layer (03:03, 06:32).
Sandboxing: isolate first, choose real-time vs batch second
Todd's product snapshots a customer's actual source code to analyze, execute, and modify it, which makes isolation non-negotiable — a leaked secret or a stray write is a real incident, not a bug report (10:38, 10:48). Rather than build container isolation themselves, they use the platform's built-in sandbox (11:28). On top of that, they split workloads by latency need: large-scale funnel verification runs as a nightly batch, and user-facing queries run in real time, kept on separate paths (11:51, 12:05).
Memory split: what you index yourself, what the platform holds
Mihir's Watchtower draws a line through memory. Account memory — a given account's lifecycle and deal history — is indexed by Watchtower itself, because contamination between accounts and precise, per-account querying both demand tight control (07:01, 09:45, 18:12). Org- and user-level memory — revenue-forecast criteria, a rep's preferences, how sales fields are defined — sits in Anthropic's Managed Memory, because that logic is shared and generic across the org rather than unique per account (07:46, 10:00). The parent Watchtower agent writes code directly to fan out across roughly 500 accounts' memory in parallel, searching and rolling up results (08:21, 08:42) — a pattern the panel calls "dreaming," an unreviewed batch pass over memory that surfaces what to act on next.
The build-vs-buy test, and how they evaluate afterward
The founders converge on one question: is the harness infrastructure our core competency (12:27)? When it isn't — when the product's edge is the briefing quality or the sales roll-up, not the plumbing under it — buying managed infrastructure wins, because an early-stage startup's scarce resource is experiment speed toward product-market fit, not infrastructure upkeep (12:33, 13:05, 13:22, 14:04). The stated exceptions are narrow: extremely latency-sensitive voice ("Jarvis"-style) products that need to control the harness directly (12:45, 12:51, 33:55), and large-traffic products that need fine-grained margin control, such as batch mode's roughly 50–75 percent cost savings (29:40, 31:06). On evaluation, the panel warns that early internal dogfooding is vibes-based and overfits to the team's own usage (24:45, 25:07); the fix is interviewing real customers about their actual query intent to build segmented benchmark cohorts (26:15). Because memory changes and external state (Slack, a CRM) drifts, these are stateful systems that resist static evals (26:42, 27:07). And when migrating models, they avoid over-tuning prompts to one model family's quirks, checking instead for defenses against the new model's own failure patterns — like a tendency toward excessive modifiers or dash overuse (27:45, 28:41).
Where it lands in Jayverse
- Rabbit: the verifier-with-veto matters more here than for a briefing. An agent that can move money needs the panel's clean-context verifier pattern applied to the mandate enforcer — fail-silent means the enforcer blocks a send rather than let a degraded approval through.
- Verex: the Watchtower pattern is the market-maker ops shape. One dedicated agent per market, with a roll-up agent above it fanning out across markets for pricing and inventory summaries, is a smaller version of what Watchtower does across sales accounts.
- Auditor: the independent verifier is the Auditor role, and the rubric is the rule. "What was checked, by which rubric" is exactly the Auditor's job description; write the rubric down the way Saumil's team did before the first wrong brief goes out.
- Dark Horse: Managed Agents is a legitimate buy option, not a compromise. For a candidate whose edge isn't the harness — most of them — building on Managed Agents keeps the repo small and the runway spent on the product decision, not on sandboxing.
- Eng: practice the three-sentence answer to "build or buy the harness." State the test (does the harness define your product's value), the two named exceptions (latency-critical control, large-scale margin control), and one concrete example from Jayverse.
Verified and unverified
Verified on 2026-09-19: Anthropic offers Claude Managed Agents — server-hosted agents with a managed sandbox, memory, and tool execution — as part of the Claude platform; "LLM-as-judge with a clean context" and "graceful degradation" (failing silently rather than showing a low-confidence answer) are established patterns in production LLM systems; batch inference APIs typically discount around 50 percent versus real-time pricing. Taken from the summary and not independently checked: the founders' names and product names (Saumil/Briefs & Tasks, Mihir/Watchtower, Todd's product), the two-week build time, the roughly 500-account figure, the "dreaming" terminology, and all timestamps.
Sources: YouTube — "How founders build on Claude Managed Agents," Anthropic's Claude channel · related items: Tech #62 (agentic engineering writes the boundaries), Tech #101 (ADK voice — the low-latency build exception), Tech #102 (MLflow — evals), harness-engineering-shift-left and vercel-eve-filesystem-agents (the own-the-environment counter-position).
Key expressions
| Expression | 뜻 · 쓰이는 자리 |
|---|---|
| harness | 하네스(에이전트를 실행시키는 인프라 전체: 샌드박스·메모리·스케줄링·도구 실행) · build-vs-buy 논의의 핵심 대상. "does the harness define your product's unique value" |
| Managed Agents | Anthropic이 제공하는 서버 호스팅 관리형 에이전트 서비스 이름 · 이 글의 주제. "shipped Claude Managed Agents to production" |
| clean context | 이전 작업의 흔적이 섞이지 않은 깨끗한 컨텍스트 · 독립 검증자를 실행 에이전트와 분리할 때 쓰는 조건. "a verifier agent that runs in a clean context" |
| rubric | 채점 기준표(무엇을, 어떤 기준으로 통과시킬지 명시한 규칙) · 검증 에이전트가 출력을 평가하는 잣대. "scoring the output against a rubric" |
| fail-silent | 실패했을 때 조용히 아무것도 보여주지 않는 설계 · 틀린 답보다 침묵을 택하는 원칙. "a fail-silent / graceful-degradation choice" |
| graceful degradation | 우아한 성능 저하(품질이 기준 미달일 때 조용히 물러나는 설계) · fail-silent와 짝을 이루는 표현. 위와 같은 문맥. |
| sandbox(ing) | 샌드박스(격리된 실행 환경), 격리하는 행위 · 고객 소스 코드를 안전하게 다루기 위한 필수 장치. "isolation... non-negotiable" |
| fan out | 하나의 작업을 여러 갈래로 병렬로 퍼뜨려 처리하다 · 상위 에이전트가 수백 개 계정 메모리를 동시에 훑을 때. "fan out across roughly 500 accounts' memory" |
| roll-up | 여러 항목(거래처·마켓 등)을 하나의 요약으로 집계한 것 · Watchtower의 핵심 산출물. "a roll-up agent above it" |
| PR | Pull Request(풀 리퀘스트, 코드 변경을 제안하는 단위) · 코드 자가 치유 제품이 제출하는 결과물. "proposes a fix PR" |
| UX | User Experience(사용자 경험) · PR 제출 시 함께 검토하는 대상. "reviews UX and instrumentation" |
| PMF | Product-Market Fit(제품-시장 적합성, 제품이 시장 수요에 맞아떨어지는 상태) · 초기 스타트업이 최적화하는 목표. "toward product-market fit" |
| funnel conversion | 퍼널(단계별 유입-전환 흐름) 전환율 · 이 지표가 떨어지면 코드 분석이 트리거된다. "when funnel conversion drops" |
| dogfooding | 자사 직원이 자사 제품을 직접 써보며 검증하는 관행 · 초기 평가의 한계로 지적됨. "internal dogfooding is vibes-based" |
| overfit | 특정 데이터(팀 자신의 사용 패턴)에만 맞춰져 일반화가 안 되는 상태 · dogfooding의 부작용. "overfits to the team's own usage" |
| stateful | 상태를 유지하는(이전 상호작용·외부 변화가 다음 결과에 영향을 미치는) · 정적 eval이 잘 안 맞는 이유. "stateful systems that resist static evals" |
| batch mode | 실시간이 아니라 모아서 일괄 처리하는 방식 · 비용을 크게 아끼지만 지연이 늘어나는 트레이드오프. "batch mode's roughly 50-75 percent cost savings" |
| runway | 활주로(스타트업이 자금이 바닥나기 전까지 버틸 수 있는 기간) · build 대신 buy를 택하는 이유. "spend your runway on plumbing" |
| core competency | 핵심 경쟁력(그 회사만이 잘할 수 있는 영역) · build-vs-buy 판단의 기준 문구. "is this infrastructure our core competency" |
| plumbing | 배관(눈에 띄지 않지만 필수적인 기반 인프라를 가리키는 비유) · 하네스를 제품과 대비시키는 말. "spend your runway on plumbing instead of product" |
| veto | 거부권(승인을 막을 권한) · Rabbit의 mandate enforcer가 갖는 역할을 설명할 때. "the verifier-with-veto matters more here than for a briefing" |