Design patterns for agentic software that doesn’t exist yet.

006·25 Apr 2026·Trust · Control

Permission Timeline

Visible history of what you've approved, with per-item revoke.

Granted permissions6 / 6 live
fs.read·src/**
Repo read access · 3 days ago
shell.exec·pnpm test
Run test suite on demand · 3 days ago
fs.write·~/.ssh/config
Append host alias · 2 days ago
net.fetch·api.notion.com
Read workspace pages · yesterday
secret.read·OPENAI_API_KEY
From .env.local · today
fs.write·CHANGELOG.md
Append release notes · today
Next agent action
Sync Notion page → CHANGELOG.md
Pull last release notes, append to file.
needsnet.fetchneedsfs.write
Ready

Approvals pile up and vanish into scroll history: no way to revoke one grant without ending the session. Permission Timeline treats the audit log as a primary surface. Every approval lands as a persistent row with scope, timestamp, and the action it unlocks. Each row carries its own revoke.

005·28 Apr 2026·Cost

Cost Meter

Forecast the finish cost, not just the current spend.

Task
Refactor authentication layer
$0.0000spent
→ $0.0900 est
0 in · 0 outSession$2.3400

An agent is eight minutes into a refactor and the spend counter reads $0.38. Useful for the bill, useless for the decision in front of you: let it run, or cancel and retry cheaper. Cost Meter pairs live spend with a locked estimate at click-time so the number means something, and tracks the session running underneath. One task isn’t the cost. The drip of all of them is.

004·20 Apr 2026·Trust · Control

Trust Dial

Per-action autonomy level that adjusts with history.

Ask for risky2 / 4
Safe actions proceed automatically. Risky and critical require approval.
Recent actions
Read config.json
src/config.json
09:41:02auto
Write changelog entry
CHANGELOG.md · 3 lines
09:41:05auto
Install npm package
zod@3.22.4
09:41:09ask
Modify auth middleware
src/middleware/auth.ts
09:41:14ask
Delete migration files
db/migrations/002_*.sql · 4 files
09:41:19ask
Lint and format
src/** · 22 files
09:41:22auto
Update env variables
.env.production · 2 keys
09:41:27ask
Run test suite
jest --coverage
09:41:31ask
safemoderateriskycritical

Agent autonomy is usually a single switch: confirm every action or run fully unattended. A large refactor and a changelog edit get treated identically. The Trust Dial aims to fix that: four levels mapped to risk tiers, with the live feed showing exactly which pending actions shift between auto-approved and held for review before you commit.

003·20 Apr 2026·Intent

Reverse Prompt

Agent rewrites your request into its interpretation before acting.

Your request
>
Request
Review
Done

Agents pick a scope from vague requests without telling you what they chose. “Clean up the auth code” becomes a diff: too wide, too narrow, or just wrong. You only find out after it runs. Reverse Prompt surfaces the interpretation first. Your request is rewritten as a precise, scoped statement you can edit or reject before anything executes. You approve the interpretation, not the action.

002·20 Apr 2026·Reasoning

Reasoning Trace

Make the agent's chain of thought inspectable, step by step.

Task
Rename AuthMiddleware → AuthGuard
Press Run to step through the agent’s reasoning.
Idle

Most agent UIs collapse reasoning, tool calls, and decisions into a single reply, leaving no way to see why the agent did what it did. Reasoning Trace surfaces the full chain of thought as an ordered, expandable log. Each step is typed (THINK, CALL, DECIDE) with a one-line summary, expanding to reveal raw content. Step through manually or reveal the full sequence at once. Trust is earned by inspection.

001·17 Apr 2026·Trust · Control

Blast Radius

Before approving, see what the agent will touch.

Proposed action
RenameAuthMiddleware → AuthGuard
reach

Agents take lateral actions. One rename hits hundreds of files, breaks a colleague’s open PR, touches a sibling repo’s import. Current tools either fire the action or show a text confirmation. Blast Radius is the missing middle: a spatial view of the full perimeter, with per-item opt-out before you approve.