Why We Handed Our Agent's Infrastructure to Anthropic -- Paridhi Agarwal (Every)
Paridhi Agarwal's first bylined piece for Every (Source Code series) is the engineering deep-dive behind the Oct 6 Every Agent GA launch: why Every moved its shared Slack coworker's infrastructure onto Anthropic's Claude Managed Agents (CMA) rather than running it themselves.
The history: Every's "Plus One" experiment (May) gave every employee a personal agent running OpenClaw on a rented VM -- and turned into an operations burden (dozens of cloud machines to babysit, a /heal-bot repair command, OpenClaw's daily 4am conversation reset, tangled memory systems, manual re-logins when a bot's ChatGPT session expired). The team concluded personal "pet" agents don't scale and consolidated to one shared company agent instead -- the pivot already tracked in [[2026-05-15-every-team-agents-vs-personal-pets]]. They then tried self-hosting that single shared agent on their own servers; within a month the maintenance load had convinced even CEO Dan Shipper, who'd expected it to be manageable, to abandon the idea. They rebuilt on CMA in about a week instead.
CMA's architecture, per Anthropic's own framing: Claude plus the harness is the "brain"; the virtual computer and tools are the "hands"; the session is a persistent log that survives either crashing, billed at 8 cents/hour while running. CMA also handles context compaction, mid-task interrupts, and prompt caching. Each Slack thread gets its own session; a small, fast model triages untagged messages to decide whether @Every should reply, react, or stay silent. The agent never holds credentials directly -- Every's own servers make authenticated calls on a user's behalf and hand back only the result.
Trade-offs the piece names explicitly: (1) a self-inflicted bug where rolling deploys could orphan an agent's in-flight request mid-session, requiring a manual fix; (2) a model-level regression when Opus 5 made @Every's replies balloon from ~280 to ~880 characters, resolved only when Opus 5.5 shipped; (3) long-running Slack threads that either forced an expensive cache-rebuild or, near the context limit, returned nothing at all -- fixed by proactively restarting sessions with a recap before hitting either wall, cutting total costs 39% when retrofitted to past conversations; and (4) trade-offs CMA itself imposes with no workaround: it only runs Claude (no cross-lab model routing even when a cheaper/better-suited model exists elsewhere), and because CMA retains session history to support resumption, Every's agent is ineligible for Anthropic's zero-data-retention option -- which "some companies require." An internal architecture doc scoping a full self-hosted replacement (harness, sessions, memory, tools, sandboxes) came back estimating months of work, which is why they're staying on CMA for now.
Why this is in the vault
A named engineer's unusually candid cost/benefit account of self-hosting vs. managed agent infrastructure -- concrete bug patterns, a quantified cache-driven cost fix (39%), and an explicit compliance trade-off (no ZDR) that's directly checkable against RDCO's own agent deployments.
Mapping against Ray Data Co
Strong mapping. The most concrete connection: Every -- a team with real engineering capacity -- tried self-hosting just one shared agent and bailed within a month because the maintenance tax wasn't worth it, choosing CMA instead. That's a direct reference point for evaluating the founder's own self-hosted, always-on Mac Mini channels-agent ([[project_channels_agent_setup]]): if a well-resourced team concluded self-hosting even a single shared agent doesn't pencil out past the prototype stage, it's worth periodically re-asking whether the Mac Mini setup should stay self-hosted indefinitely or whether a CMA-style managed layer becomes worth the lock-in (single-model-family, no ZDR) as usage grows. Second, the piece hands over the exact mechanism behind the open DECISION NEEDED item on cache-breaking patterns: don't let a session idle until its cache entry expires or its context fills -- proactively end the session with a short recap before either happens. That's a directly portable technique, not just a data point, and it complements [[2026-10-07-every-efficient-agent-token-caching]]'s finding that silent per-user identifiers in a system prompt can defeat caching with zero error signal -- together the two pieces describe both halves of a cache-hygiene checklist (don't vary the prompt per-user; don't let sessions run past the cache or context horizon) worth auditing against the Mac Mini agent directly. Third, the no-ZDR disclosure is a concrete fact to know before ever positioning a CMA/Claude-Code-based architecture to a phData-style enterprise prospect with data-retention requirements.
Related
- [[2026-10-06-every-agent-launch-shared-slack-coworker]]
- [[2026-10-07-every-efficient-agent-token-caching]]
- [[2026-05-15-every-team-agents-vs-personal-pets]]
- [[project_channels_agent_setup]]
⚠️ Sponsorship
No third-party paid block. But this is Every writing about its own infrastructure choice in service of selling its own Every Agent product (launched two days prior) -- a structural self-interest angle even without a disclosed sponsor. The standard house footer also runs underneath (Every All Access subscription pitch, Sparkle/Cora/Spiral/Monologue cross-promo bundle). Read the trade-offs section as unusually candid for a vendor post, but the overall framing ("staying on CMA lets us spend our time on @Every instead of infrastructure") still serves Every's narrative that its own product is worth paying for.