Forecast // April 15, 2026
Built on the assumption that Anthropic develops Claude Code with a Ralph-Wiggum style autonomous loop — a model that finishes the backlog, then asks itself what's missing and writes the next round. If that's true, the backlog is predictable: it's whatever the shipped docs already admit is broken.
Anthropic shipped roughly 80 distinct Claude Code features between Feb 1 and Mar 24, 2026 — on the order of 1.5 per day. That cadence is not human-paced engineering planning. It is what happens when a harness keeps firing, the backlog is regenerated from "what does the product still lack," and every shipped feature feeds the next round of ideas.
Context Foundry uses the same pattern. After the task queue empties, a discovery agent scans the built state and appends new tasks. The loop is self-extending. If Anthropic's loop is smarter (almost certainly), the adjacencies it walks are still legible from outside: read the docs and look for the phrases "not yet supported," "individual only," "research preview," "cannot currently." Those are the loop's next inputs.
This forecast went through two rounds. Round one leaned on pattern extrapolation from a shipped-features calendar plus some leaked feature-flag names (KAIROS, BUDDY, Agent Swarms). Round two — after an auditor pass — threw out the leak-based bets and anchored everything in primary sources: the routines documentation, the desktop redesign blog post, the Managed Agents engineering writeup, and the Claude Code changelog.
The strongest predictions are the ones that quote documented gaps verbatim. If the docs say "routines belong to your individual claude.ai account and are not shared," then "team-shared routines" isn't a prediction — it's a pre-announcement.
Each of these is a direct read of explicit "not yet" or "individual only" language in docs that shipped within 48 hours of this forecast. About as safe as roadmap prediction gets without actual insider info.
High confidence (80%+)
The routines doc explicitly states the CLI cannot currently create or revoke tokens, and GitHub triggers are web-UI only. Gaps called out this plainly are on the sprint board.
Routines are live but scoped to an individual account. Obvious next layer for Team and Enterprise plans.
Currently every matching event spawns a fresh session. Doc calls this out as a known limitation.
Unattended autonomous jobs need ops controls. Research preview stops being preview when this lands.
Long-running runs need postmortems. Anthropic keeps publishing about harness visibility.
/fire API trigger 82%Today JSON bodies are passed as literal strings. The dated beta header experimental-cc-routine-2026-04-01 implies schema iteration.
Only pull_request and release today. push, workflow_run, check_run, and issues are glaring absences for alert-triage and docs-drift use cases.
Anthropic is already packaging repeatable workflows; reusable templates are the natural follow-up.
The Advisor tool launched April 9. Productizing it inside Claude Code is a short hop.
Medium-high (70-79%)
Cowork already has RBAC, spend limits, analytics, and OpenTelemetry. Claude Code automation will inherit the same.
Current minimum is one hour. First thing power users push against for real-time alert triage.
Already exist at a basic level; finer-grained event controls are a likely iteration.
Docs already expose branch-push restrictions, which usually expand into richer policy settings.
Doc states "no permission-mode picker and no approval prompts during a run." Enterprise cannot ship without an escape hatch.
The April 14 redesign is phase one for parallel work. Bulk pause, archive, kill, resume fit the trajectory.
Anthropic is leaning into "many things in flight." Orchestration views are the logical next piece.
The /fire API is preview behind a dated beta header. Clean stable interface is the usual next step.
Today "all connectors included by default" is a placeholder. Reusable bundles or per-routine profiles are the fix.
Routines bundle repos, prompts, environments, and connectors. Org-level reuse is the missing layer.
The Managed Agents post explicitly frames private resource access as a real customer ask.
Local, cloud, remote, and desktop surfaces are all live. Smoother handoff is consistent with the architecture.
Medium (60-69%)
Anthropic's own writing on parallel agents acknowledges coordination pain. Product work usually follows.
PR review, deploy verification, docs drift, alert triage are already documented use cases.
Routines already support multiple repos. Dependency-aware execution is a plausible next step.
Anthropic already ships persistent phone control for Cowork and channels-based approval relay in Code.
Round one had KAIROS, BUDDY, and Agent Swarms as named bets sourced from leaked feature flags. The auditor was right to cut them: the functional versions of each (longer autonomy, pair mode, parallel agents) are already live under different names, so quoting codenames was implying insider knowledge that wasn't there. Also dropped: 2M context, fine-tune-from-session-history, on-device Haiku, Figma bi-directional edit, Linux desktop — all speculation with no primary-source signal.
The meta-lesson. When a product ships a research preview, the docs are the roadmap. Mining "not yet supported" phrases beats any amount of pattern extrapolation from a shipped-features calendar.