This is the "agents.vxs.ai" view — every agent that has a name and a job, in one place.
Sara
Ops Sentinel — Jira chase/nudge engine, escalation, fire alerts
LIVE since 14-Jul-2026
DigitalOcean droplet
Scans stuck Jira items on a schedule, nudges the owner directly by email (no human drafts or sends it), tracks follow-ups, and escalates to Apurva after 3 unanswered nudges or no resolvable owner. First live run: 207 real nudges, 0 errors.
Runs on: standalone droplet (139.59.24.166), not Cowork
Owner: Apurva Mankad
⚠️ Not in webxpress-servers inventory yet. Not visible/pauseable from Cowork — needs SSH to change.
Mira
Delivery Ops — daily brief, weekly PM scorecard, ETA chasers
Active daily
No closed loop as of 11-Sep-2026
Fully automated report generation and send — daily project brief (~05:30 IST), weekly PM scorecard, event-driven Jira ETA chaser emails. Update 11-Sep-2026: the Devang Mail Autopilot Routine (found 08-Aug-2026, genuinely closed-loop) and a separate, previously-undocumented Harshad routine were both intentionally decommissioned by Apurva. Mira currently has no confirmed closed-loop automation — everything she does is a report a human still has to read and act on.
New 17/18-Sep-2026 — LSP EDD/POD chaser (ABLBL): a weekly job that reads Control Tower, finds shipments past their EDD that are still undelivered and delivered shipments with no POD, builds one Excel per logistics partner and emails it as Mira. It runs from cron on the Hawkeye production server (Mon 10:00 IST) — the first agent job outside the 139.59.24.166 droplet and outside Cowork. Recipients come from a reviewed LSP contact workbook held on the same server. It is in preview (all mail to the project team + Apurva) until the team signs off those contacts; going live makes it Mira's first real work loop — chasing a third party, not reporting to us.
Sends as: Mira@webxpress.in
Underlying: customer-project-tracker skill; LSP chaser is standalone Python on Hawkeye
Also scheduled: ABLBL Weekly LogiCloud Order-Health Digest (Cowork, Mon 11:00 IST)
✅ Skill install gap resolved (confirmed enabled 08-Aug-2026). ⚠️ Still no GitHub repo of her own — the LSP chaser lives only on the Hawkeye server, with a copy on OneDrive, not in git. ⚠️ Devang Autopilot and the Harshad routine stay decommissioned (removed 11-Sep-2026). ⚠️ Two ABLBL jobs now run an hour apart every Monday (chaser 10:00, order-health digest 11:00) over overlapping data — worth merging.
Sharon
Marketing & Sales — campaign drafting/send, daily sales brief, weekly digest
Campaign send: human-gated
Briefs/digest: L3, no review
Drafts the full campaign (brief → HTML email → Brevo send → stats digest) but every campaign send is gated on Apurva's approval. New (found 08-Aug-2026): two real Cowork Routines run with zero human review — a Daily Sales Brief (08:30 IST) and Weekly Marketing Digest (Mon 08:00 IST), both sent via Brevo with a Microsoft Graph fallback if Brevo credit runs out. Also new: a droplet-hosted Sales Funnel Sync (07:00 IST daily) — the actual "sales lead sync" — matches mail to open deals in Postgres and writes safe updates directly to the database, flagging anything risky as a To Do instead. This is the most autonomous write-capable job in the whole fleet.
Sends via: Brevo
Skills: sharon, send-mail, apurva-voice, theme-factory, sales-management
⚠️ Daily Sales Brief went dark for 17 days in Jul-2026 (silent failure) before the fallback was added. ⚠️ Sales Funnel Sync's marketing.deals table and the sales-management skill's Google Sheets tracker are confirmed separate, unreconciled data stores — real pipeline fragmentation, not just a doc gap.
Baldev
Control Tower — chases drivers/transporters on HawkEye events
Not confirmed live
Proposed voice/call layer for HawkEye: vehicle delayed, checkpoint missed, driver unreachable → Baldev calls/messages the relevant party. This is the one agent that would replace a human phone call, not just a nudge.
Repo: hawkeye-codeagent (flagged "production structurally unreachable")
⚠️ Open questions: is he actually calling anyone today? Voice, WhatsApp, or both? No Twilio caller ID confirmed. This is the biggest "is it real or paper" gap in the ecosystem.
Honest classification of what each agent function does today. Level 3 = genuinely closes the loop with zero human touch, end to end. Most of the ecosystem sits at Level 1–2: it tells a human what to do, faster and more reliably than before — which is real value, but it is not headcount-replacing work yet.
1. Take the LSP chaser live, then let it close its own loop
It already builds the lists and the mails without help; the only thing missing is permission to send them to the logistics partners. Once live, the next step is a reply-reader: the partner answers with a delivery date or a POD, the agent files it against the shipment and stops chasing that one. That is a real loop — the agent handles the exchange end to end and a person only sees what is still unresolved. Target: live 28-Sep-2026, once the contact workbook is signed off.
Why this first: it is the fleet's clearest step from "sends reports" to "gets a third party to do something", and it is one setting away.
2. Fix ABLBL's EDD and POD data at source, not in the report
The first run found about 24,000 open shipments that cannot be judged at all (no EDD, and no lane TAT for that partner — Speedex has none loaded), about 2,300 carrying an EDD on or before their booking date, and a booking feed that has been silent since 11-Sep. The chaser works around all three, but every workaround weakens the mail we send a partner. Load Speedex's TATs, fix the EDD at source and restore the feed.
Why: an agent chasing someone else's SLA has to be right, or it loses the argument the first time a partner checks.
3. Confirm and productionize Baldev, or shelve it
This is the only agent designed to replace a human action a person is doing right now (a phone call to a driver), rather than speeding up a report. Right now it's unclear if it's live at all. Worth a one-week push to get a real answer: is hawkeye-codeagent deployed, does it place a call or send WhatsApp, and on what trigger volume. If it's real, it's the highest-leverage agent in the fleet. If it's not, stop counting it as "done."
Why this first: it's the one place the roadmap already points at full automation — it just needs verification, not a new idea.
4. Let Sara auto-close, not just auto-nudge, on defined ticket types
Sara already nudges and escalates without review (L3). The next step is narrow but real: for a specific class of stuck items — e.g. "spec acknowledged, no update in 5 days, low severity" — let her auto-transition status or auto-close with a logged reason, instead of only nudging forever. Start with one non-critical ticket type as a pilot.
Why: it's a small extension of infrastructure that already works live, not a new build.
5. Tier-1 customer support answering, not just triage
Sara currently scores and routes support email/Teams threads but a person still answers every one. The WebXpress Wiki connector (43k files, semantic search) already exists. A pilot agent that auto-answers the most common, lowest-risk question categories (status checks, "how do I," known-issue lookups) end-to-end — with a hard fallback to a human for anything outside a confidence threshold — would be the first agent doing real customer-facing work without a person in the loop.
Why: support is the highest-volume repetitive human task visible in the ecosystem right now, and the knowledge base to answer from already exists.
6. Auto-send low-risk marketing sends, keep approval for the rest
Sharon's approval gate makes sense for new campaigns but is overkill for repeat, low-risk sends (e.g. a recurring newsletter with a fixed template, or a re-send to a segment that's already approved once). Tiering the approval gate by risk — auto-send for known-safe patterns, human approval only for new content or new audiences — would remove a human step without changing where judgment is actually needed.
Why: cheapest possible win — no new capability, just a policy change on an agent that already exists.
7. Bring Mira's Daily Brief/Scorecard fully under Cowork so it's governable — and decide what replaces the Devang/Harshad automations
It was achievable: the Devang Mail Autopilot was a registered Cowork Routine (found 08-Aug-2026, visible and pauseable) before Apurva removed it 11-Sep-2026, along with a Harshad routine that was never documented here to begin with. Mira's older Daily Brief/Weekly Scorecard schedule still isn't a registered Routine either — same agent, same identity, inconsistent governability, and now zero closed loops instead of one. Worth deciding explicitly: was the removal a permanent scope pullback, or a placeholder until something replaces it? Either answer should get written down, not left as a gap the next audit has to rediscover. Sara still lives entirely outside Cowork on a droplet reachable only by SSH — same underlying issue, harder to fix.
Why: this isn't a new automation, it's the visibility layer this whole exercise (this dashboard) is trying to establish — without it, "how far have we progressed" stays a guess, and can go backwards unnoticed.
8. Assign an owner to "DO Server Health Check," or fold it into Sara's remit
Found 08-Aug-2026: a daily 09:07 IST Cowork Routine monitors the WebXpressTMS droplet (www.webxpress.ai / sitestaging) — real, live, but not attributed to any of the 4 registered agents. It doesn't fit Sara (Jira/support), Mira (delivery ops), Sharon (marketing/sales), or Baldev (Control Tower). Either give it a name in the registry or explicitly fold it under Sara's ops-sentinel remit.
Why: an unowned automation is exactly the kind of thing this dashboard exists to catch — it was invisible until this audit.
9. Resolved: "sales lead sync" was a droplet cron, not a Cowork Routine — and it's the fleet's best example of real automation
Found by grepping crontab -l directly on the droplet (it doesn't show up in list_triggers because it's not Cowork-hosted). Sales Funnel Sync matches Apurva's mail to open deals and writes safe updates straight to a Postgres marketing.deals table, escalating only what's uncertain. Worth studying as the template for extending similar "write directly, escalate the risky part" logic to other agents — e.g. Sara's ticket auto-close idea in #2 above.
Why this matters more than the other findings: everything else in this audit was a report running unattended. This is the first confirmed case of an agent doing the underlying work itself.
10. Reconcile the sales pipeline — two systems of record, confirmed not talking to each other
Sales Funnel Sync writes to Postgres marketing.deals. The sales-management skill's Daily Sales Brief reads from a separate Google Sheets tracker. Apurva confirmed 08-Aug-2026 these are genuinely separate and unreconciled — the same pipeline has two different sources of truth today.
Why: this is a direct instance of the "repo of repos" consolidation problem this whole project exists to fix — not a hypothetical, a live one found while building this dashboard.
11. Give the 9 Velocity refresh jobs an owner and a page
The claude-runner user on the 139.59.24.166 droplet runs nine cron jobs — Velocity report refreshes, a nightly backstop re-export and a retry — that appear in no agent's definition and on no page until today. They are doing real work for customer reporting. Either attach them to an agent with a name, or register them as infrastructure with an owner, a log check and an alert when they fail.
Why: they were found the same way "sales lead sync" was — by reading a crontab. Anything unowned stays invisible until it breaks.
12. Close the Hawkeye server exposure before more agents run there
The Hawkeye production server now hosts an agent job, and a review on 17-Sep found its database, Redis and app ports open to the internet, the main database login accepted from any IP, and every service running as root (Jira SOC-128, with Harshad). Agent jobs will keep landing on this box; the base it runs on should be closed first.
Why: the fleet is spreading to a third server. Each one inherits whatever posture it already had.