A shared review surface
Approval happens where the team can see it, not in one engineer's terminal.
Invite-only private beta · review layer for coding agentsApproving 93% of prompts
Approving 93% of prompts
isn't review.
Caskade moves the review upstream: review intent before execution. Scope, blast radius, validation, and stop conditions get approved at plan time — before the repo is touched.
For teams already running coding agents on real repos and feeling approval fatigue set in.
Goal Contract · gc-0142DraftUnder reviewApproved
outcome: add rate limiting to /api/export
blast_radius:
allow: src/app/api/export/** · src/lib/rate-limit.ts
deny: src/lib/auth/** · migrations/**
validation: bun run test · bun run build
stop_on: scope_drift · failing_checks · denied_path_write
▸ agent executing inside boundary
✎ writesrc/lib/rate-limit.tsok
✎ writesrc/lib/auth/session.tsBlocked
⤷ outside blast_radius — returned to review
✓ handoff:reviewed change + evidence packet
93%
of permission prompts get approved by Claude Code users — Anthropic's own number.
Anthropic engineering ↗In Anthropic's own eval, its dangerous-action classifier missed 17% of genuinely overeager actions.
Published auto-mode evaluation ↗“The user becomes a rubber stamp.” Approval that happens mid-run decays into reflex.
Runtime prompts aren't governance
A permission prompt is not a review.
Provider guardrails ask one question, mid-run: allow this command? They can't ask whether this is the change your team agreed to make. That question only makes sense before the run, against the whole plan — transcripts and diffs answer it even later.
An independent stress test of the permission gate found its main failure mode: it hunts for approval-shaped evidence in the conversation without checking whether that approval covers the actual blast radius of the action.
Independent stress test · arXiv:2604.04978 ↗
Runtime guardrails are the last line of defense. Keep them. They were never designed to be the first.
- Review the plan while it can still change
- Approve a blast radius, not a command
- Escalate drift back to review, not to a prompt
- Hand off a change with evidence
Standard workflowReactive
Agent mutates repo
Transcript archaeology
Diff cleanup and fixes
Caskade workflowProactive
1Goal Contract drafted
2Human reviews intent and bounds
3Agent runs inside blast radius
Plan mode is a step, not a surface
A generated plan is not a reviewable artifact.
Plan mode was the right instinct: look before the agent leaps. But generated plans run long, repeat themselves, and live in one engineer's terminal. A plan that gets skimmed puts you back to trusting the model by default.
A reviewable plan is short, structured, and binding: an outcome, a verification surface, constraints, a blast radius, an iteration policy, and stop conditions. Something a second engineer can edit and approve — not scroll past.
Verbose isn't reviewable.
"The new plan mode is too verbose… one conversation already used 42% of my 5-hour quota."
The loop
A tighter loop for real agent work.
The plan becomes a Goal Contract — outcome, verification surface, constraints, blast radius, iteration policy, stop conditions — reviewed by your team before the repo is touched.
01
Goal Contract
The work starts as an explicit outcome, verification surface, constraints, blast radius, iteration policy, and stop conditions.
02
Intent review
Humans review and edit the contract before execution proceeds.
03
Bounded execution
The agent run stays inside the approved blast radius.
04
Drift checks
Meaningful scope changes escalate back to review — they don't become another mid-run prompt to click through.
05
Clean handoff
The outcome is a reviewed change with evidence, not transcript archaeology.
Reviewed work compounds
Transcripts disappear. Reviewed intent compounds.
Governance isn't only about blocking bad runs. A reviewed contract is a shared artifact: the team sees what was approved, why drift came back, and what evidence shipped — and the review habits worth standardizing emerge from that record.
Drift is a decision, not a surprise
When a run escalates, the whole team sees what changed and who re-approved.
Evidence outlives the run
Handoffs carry the contract and its verification, not a transcript to excavate.
Contracts become house style
Reviewed intent teaches the next contract. Hidden transcripts don't.
Who it is for
Built for teams already letting coding agents touch real repos.
Caskade is for founder-engineers, small AI-native engineering teams, and technical leads already using Claude Code, Cursor, OpenCode, Codex-style tools, or similar workflows in live repos.
These teams do not need another pitch for agent upside. They need a cleaner way to decide what agents are allowed to do, when drift needs review, and whether the finished work matches the original intent.
It is also for the platform, security, and DevOps leads being asked how agent changes are governed — who today have nothing upstream of a permission prompt to point at.
Caskade does not compete with the agent.
It is not another IDE agent, generic copilot, or CI replacement. It doesn't compete with your provider's runtime guardrails either — it's the review that happens before they ever fire.
It competes with the messy gap between an agent saying "done" and a change you can trust.
Invite-only beta for teams already using coding agents on real work
Bring us a real workflow.
We are opening Caskade to a small first wave of beta testers already using coding agents on live repos.
- Bring a real repo, workflow, or spec
- Help shape Caskade around real review pain
- Get a private beta invite if the current wave is a fit
What happens if you apply.
- 1
You leave your email and share the kind of coding-agent workflow you want reviewed.
- 2
We review fit for the current invite-only private beta wave.
- 3
If it is a fit, we send a private app link and work with you on a tighter intent-review loop.
Governance starts before execution,
not at the prompt.
Caskade is that upstream review layer. The app is invite-only for now; leave your email if you want to be considered for the first beta wave.