About Stride

What Stride is, how it works, and what it demonstrates.

Stride is a multi-agent orchestration tool that does the work a project manager otherwise spends a morning on — synthesized status, surfaced risk, prepped meetings, audience-tailored stakeholder comms, and a curated project tracker — with human review at every handoff and before every external action.

01

Three workflows

Workflow A

Morning Brief

Pulls from your sources in parallel, drafts your standup and exec update, surfaces today's meetings with prep, and reconciles the tracker. One synthesized brief in minutes.

Workflow B

Meeting → Artifacts

Drop in a transcript. Get action items with owners, three audience-tailored summaries (team / exec / client), risk register entries, and ticket drafts.

Workflow C

Status Tracker

A persistent project view that updates with every approved artifact: RAG status per workstream, milestone timeline, open risks, action item board, browsable history.

02

Agent roster

Seven specialized agents plus one shared reviewer. The orchestrator plans which doers run and in what order; the reviewer critiques every doer's output before a human ever sees it.

  • Orchestratorplanner

    Plans the run, dispatches doers, assembles the final brief, archives to history.

  • Status Synthesizerdoer

    Sprint board + threads + mail → factual status read with velocity, slipping, and on-track items.

  • Risk Detectivedoer

    Surfaces blockers, slips, and dependencies; deduplicates against tracker; flags risks aging out.

  • Meeting Prepdoer

    For each meeting today, an agenda + prior context + 3-5 actionable talking points.

  • Comms Tailordoer

    One status, four audience drafts: standup, exec, client, skip-level. Calibrated tone per audience.

  • Action Trackerdoer

    Extracts action items from transcripts and threads; reconciles against existing tracker items.

  • Tracker Curatorcurator

    Reconciles approved outputs into persistent state; recomputes RAG; appends history; saves.

  • Reviewer / Analystreviewer

    Critiques every doer output before a human reviews it. Verdict: pass / flag / block, with confidence.

03

Human-in-the-loop, by design

Four structural gate types make sure no agent ever acts unilaterally:

  • Between-agent handoff

    After every doer + reviewer pair finishes, the user can edit the output before it feeds the next agent. Edits flow through the gate and become the next agent's input.

  • External action

    Sending email, posting Slack, creating Jira tickets, scheduling calendar events — each fires an explicit approve / edit / reject card. Nothing leaves Stride without a click.

  • Final publish

    Morning brief and meeting artifacts always end with a final review-and-publish gate. The tracker only commits on approval.

  • Low-confidence escalation

    Reviewer can flag any output as low-confidence, which forces a mandatory human pause even in a fast-path mode. Escalation is structural, not optional.

On every gate, the payload renders as a real card (not JSON). The PM can either edit inline (e.g. calendar events have date / time / duration / attendees fields, plus include / exclude toggles) or type a natural-language instruction (“move all to 1 hour, drop the Tuesday block”) and a Claude call rewrites the payload in the same schema.

04

Architecture decisions worth highlighting

  • Doer + reviewer pattern. Every agent output passes through a separate reviewer agent before reaching the human gate. Mirrors how good PMs work: do, peer-check, then ship.
  • Integration registry. Adapters self-register based on env vars; agents query the registry, never specific adapters. Adding Jira, Linear, Slack, or Notion is a one-file drop with no agent or orchestrator changes.
  • SSE-first orchestration. Agent events stream over Server-Sent Events; the trace panel renders state as it happens. HITL gates are open Promises that resolve when the user clicks approve.
  • Eval rigor. Per-agent Zod schemas enforce shape correctness; golden sets test “must mention” / “must not mention” assertions per audience. Run shape-only in CI, live with model calls when changing prompts.
  • Corporate-formatted exports. Every workflow output extracts to a Word doc with proper page header / footer, document properties, data tables, and Microsoft-standard typography. Calendar exports are interactive, gated, and idempotent.
  • Brand discipline. One palette, two fonts, sharp corners, no emojis. Same visual language as the rest of the portfolio (Attribution Truth-Checker, Triage, Pack).
05

What this project demonstrates

  • Multi-agent orchestration design

    Decomposing a fuzzy job into specialized agents, wiring them with an orchestrator, and adding a verification layer that critiques every output.

  • Human-in-the-loop maturity

    Every gate has a clear surface, a way to revise (natural language or inline), and a default-to-skeptical posture.

  • PM domain depth

    Workflows calibrated to actual PM rituals — RAG status, ship-date confidence, attorney sign-off bottlenecks, exec readout prep.

  • Production architecture

    Registry-based extensibility, SSE streaming, persistent state across hot-reload, env-gated adapters. Production decisions, not toy decisions.

  • Evaluation craftsmanship

    Zod schemas + golden sets + shape and live modes. Cost-aware testing that catches regressions before they ship.

  • Output craftsmanship

    Drafts read like prose. Word exports use real corporate formatting. Calendar exports are interactive and gated.

Try it

Run a morning briefTry meeting modeOpen the status tracker

The synthetic project — Northridge Financial Phase 2 — is modeled on a real CLM AI delivery engagement. Same workstreams, same risks, same accuracy gates that Chantel runs at her day job.