System & reliability brief
Meetings end. Follow-through is verified.
Delegate.ai is one agent with one job: turn what was said in a meeting into what actually happens next. It parses transcripts, plans the work, holds for your review with the full queue on screen, executes across Google Calendar, Gmail, Slack and Notion, and proves what landed — with a verification receipt on every step and an evaluation suite that runs live.
How the agent works
1 · Intake agent
A transcript — pasted into the dashboard or captured live by an AI notetaker bot — is parsed into structured actions: task, assignee, deadline, priority, and the app that should handle it.
2 · Planner agent
Each action becomes an executor-shaped plan: ordered steps with concrete params, one app each. A single action may chain steps across apps (schedule it, then announce it) and address the resource a previous step created.
3 · Executor agent
Deterministic TypeScript executors — not the LLM — perform the writes: Google Calendar events, Gmail drafts, Slack messages, Notion blocks. The model plans; it never touches an API directly.
4 · Verification receipts
Every created resource is read back or acknowledged before the step is reported as executed: calendar events are re-fetched by id, Notion returns the block ids it created, Slack answers with its literal ok. Anything else is a skip or a failure with the reason on the step.
Reliability model
Honest failure semantics
A step whose app is not connected is reported as skipped — with the app named and the fix in the detail — never as a success. Partial success is the normal case with real APIs, and the run panel reports exactly which steps landed.
Nothing is fabricated
Email recipients are never guessed: a draft without a real recipient from the transcript is held, and the executor validates recipient shape before composing. Calendar events require a resolvable date — “sometime next week” is skipped, not invented. Template tokens that reference a step which did not complete resolve to an explicit unavailable marker.
Provider fallback chain
Parsing and planning run through a model chain (Gemini primary, a second provider, then a deterministic local parser) so a missing key or exhausted quota degrades to the honest local-heuristic label instead of a hard failure.
Bounded transient retry
All three connector call layers retry once with backoff on clearly transient responses (408/429/5xx, network resets). Exhaustion surfaces as the last real error; non-transient answers never retry.
Measured, not claimed
The committed baseline measures extraction quality on a golden dataset (micro-F1 0.875, 6 cases, gemini-3.7-flash). The in-app suite at /evals goes further: it runs the live pipeline against routing, precision, priority-escalation and no-fabrication scenarios on demand, plus deterministic executor proofs — every check's evidence is displayed.
Safety defaults
The run parks at review
Parsing never touches an app. Every run holds at a review gate with the full queue on screen — email delivered as draft by default, real sending an explicit opt-in, and nothing executes until you press Run.
Least-privilege connections
Google: calendar.events + gmail.compose only — never full-calendar or full-account scopes. Slack: a webhook that posts to exactly one channel and reads nothing. Notion: an internal integration that can touch only the shared page, append-only.
Credentials never reach the page
Every app credential — the Google token, the Slack webhook, the Notion token — is sealed in an AES-256-GCM HttpOnly cookie bound to one account, used server-side, and never written to a database.
Append-only knowledge writes
The agent can add a Notion record of what it did; it cannot rewrite the user's page.
Where each scored criterion is proven
The rubric, answered in place — every criterion links to the surface where you can verify it right now, signed in or not.
- 30%
Technical execution
Four real executors — Google Calendar, Gmail, Slack, Notion — driven by multi-step plans that chain across apps and address resources a previous step created. Credentials are AES-256-GCM sealed in HttpOnly cookies; connector calls retry transient failures with backoff.
prove it: /demo run · src/lib/google-workspace.ts
- 25%
Reliability & evaluation
A live 9-scenario suite with per-check evidence, persisted run history, honest skip/fail semantics on every step, a measured 0.875 micro-F1 baseline on a golden dataset — and a review gate that parks every run until you press Run, so email is never sent and Calendar is never written on parse alone.
prove it: /evals · the review pause on any dashboard run
- 20%
Usefulness
Meeting ends → follow-through executes in the apps the team already uses: events booked, recaps composed, channels posted, decisions logged. Passwordless sign-in and paste-a-webhook connectors make the first five minutes real.
prove it: /demo · /signup magic link
- 15%
Originality
The honesty contract: nothing is fabricated — every executed step carries a provider-acknowledged receipt, every gap says why, and the eval scorecard is a product surface rather than a footnote.
prove it: receipts on any run · this page
- 10%
Demo clarity
A one-click guided showcase that runs the real pipeline deterministically, the two-minute script below, and receipts that narrate themselves — the demo explains itself while it plays.
prove it: /demo · script below
Two-minute demo script
The recorded demo follows these beats — the same flow a judge can drive live in under two minutes.
- 0:00
Landing page
One sentence: meetings end, follow-through is automatic. Delegate.ai turns meeting notes into verified action across Google, Slack and Notion.
- 0:20
Guided showcase (/demo)
One button runs a curated four-app transcript through the real pipeline — no typing on camera. Prefer the raw flow? Paste your own transcript in the workspace instead.
- 0:40
Watch the agent run
The run panel narrates: transcript analyzed → plans generated → the run pauses at review, where you choose email delivery — draft or real send — with the parsed email steps on screen. Press Run and point at the verified badges: every step is acknowledged by the provider, and the queue shows the honest detail per action.
- 1:05
Connected Apps
Show the four apps. Connect Slack with a paste-credential webhook and send the live test message to the channel — a real, visible delivery.
- 1:25
Evals — the proof
Press Run the suite on /evals: the live scorecard reruns routing, precision and no-fabrication scenarios against the real pipeline and shows 100% with per-check evidence. Run history below proves it holds over time.
- 1:45
Reliability recap
Close on the philosophy: the agent reports what actually landed — verified receipts, honest skips, partial failures visible per step. Judges can curl /api/evals themselves.