thirdshift — the shift nobody staffs

Your CMMS knows that pump 3 is down. Nothing in your stack gets the part on the bench. thirdshift is the night clerk that does.

Last Mile Agent Hackathon, AWS Builder Loft SF, 2026-07-21. Novita (inference) + ActionLayer (browser execution). Numbers below are either measured live tonight with the ticket ID attached, or cited in RESEARCH.md. python3 verify.py re-checks every ticket against the API.

The problem

Unplanned downtime is the most expensive event in a plant. The world's 500 largest manufacturers lose $1.4 trillion a year to it, 11% of revenue, and an idle automotive line runs $2.3M an hour (Siemens/Senseye, "True Cost of Downtime 2024").

Between "machine is down" and "part is ordered" sits one person, translating a symptom into a catalog specification:

"Bearing on pump 3 is squealing" → 6203-2RS, double rubber-sealed, 17 mm bore, 40 mm OD, 12 mm width, qty 2.

That translation is tribal knowledge, and it is retiring. US manufacturing needs 3.8M new workers by 2033 and expects 1.9M positions to go unfilled, with Baby Boomer retirement a named driver (Deloitte + The Manufacturing Institute, 2024). Small facilities never had that person to begin with. In a property portfolio, a single-line food plant, or a machine shop, the maintenance manager is procurement.

Nothing in the stack closes the gap:

• CMMS stops at paperwork. MaintainX's own feature pages cover low-stock alerts, PO generation, and vendor records. A human still identifies the part and places the supplier order.
• The catalog has no door for small buyers. McMaster-Carr's data API is approval-gated and returns product data only, no ordering. Punchout and cXML assume you already run e-procurement. At ten employees, you get the browser.
• MRO is unmanaged tail spend. A $700–770B market in which roughly 80% of purchase transactions account for 20% of spend, and nobody is watching them.

Getting the translation wrong costs twice. In adjacent industries about one in five parts ordered online comes back, and an avoidable truck roll runs $150–$500. The number that matters is work-order-open → part-on-bench, and every hour of it on a down line is the most expensive hour in the plant.

What thirdshift does

Type what the technician said. thirdshift asks the two to four questions the retiring senior tech would have asked, compiles one exact imperative goal, and then puts a real browser on the supplier and utility portals overnight — the sites that will never give a small facility an API.

One queue, four night-clerk moves:

mode	what the night clerk does	stops at
plant.py "…"	exact part on a distributor catalog: part №, price, stock	read-only
--cart	filled cart + what checkout requires to order	before "Place Order"
--warranty	manufacturer RMA claim, completed from the work order	before final submit
--rebate	utility rebate owed for the efficiency swap: program, amount, deadline	read-only

Every mode stops before the irreversible click. The human signs; the night clerk did everything up to the signature.

No supplier is hardcoded. --supplier mcmaster|grainger|zoro|motion repoints the same work order at a different catalog, and the compiler adapts the goal to it — "Grainger item number" against grainger.com, "McMaster-Carr part number" against mcmaster.com, from one work order and one code path. The recovery modes go further and take no target at all: Novita resolves the manufacturer's support site and the utility's rebate portal from the facts, which is how Grundfos and PG&E came out of the same code without either being named in it. Try all four in about four seconds each:

python3 plant.py "bearing on pump 3 is squealing" --supplier grainger --facts workorder.json --dry
python3 plant.py "bearing on pump 3 is squealing" --supplier zoro     --facts workorder.json --dry

The recovery modes go after money the facility already owns and never collects, because a portal form stands in front of it. Nobody files an RMA for a $40 bearing, so a year of work orders quietly donates real money to vendors. Utility rebates expire on deadlines nobody tracks. Compliance renewals carry fines that dwarf the filing effort. thirdshift is already holding the part, the failure date, and the symptom, so the claim writes itself.

One night shift, drawn to scale

gantt
    title 22:00 → 00:00 — six items, zero humans awake
    dateFormat HH:mm
    axisFormat %H:%M
    section parts
    WO-114 source bearing (McMaster)      :a1, 22:00, 20m
    WO-115 source drive belt (McMaster)   :a2, after a1, 20m
    WO-117 fill cart, qty 2 shaft seals   :a3, after a2, 20m
    section recovery
    Grundfos warranty claim (WO-102)      :b1, after a3, 20m
    PG&E motor rebate lookup (WO-102)     :b2, after b1, 20m
    Boiler cert renewal, state portal     :b3, after b2, 20m

Loading

First shift arrives to part numbers, a filled cart, a completed claim form, and a rebate deadline, none of which existed at close of business.

The money — one plant, one year

Reference plant: single line, ~$20M revenue, ~20 sourced-part work orders a month, one maintenance manager doubling as procurement. Inputs and sources are in RESEARCH.md §7–10, and where a source gave a range the model takes the low end.

Money saved:

line	mechanism	year one
down-line hours removed	the part gets sourced overnight instead of whenever the manager gets to it; counting only 4 events a year where the part path is critical, 1 hour each, at the $10K/hr discrete-manufacturing floor	$40,000
wrong-part reorders avoided	about 1 in 5 parts ordered online is returned; an exact spec halves it — 24 avoided reorders a year, rush freight plus repeat labor	$4,000
manager hours returned	240 sourcing sessions a year × 30 min × $58/hr (BLS median, industrial production manager)	$7,000

Money made — cash in, not cost avoided:

line	mechanism	year one
warranty claims filed	manual tracking leaves most eligible claims unfiled; systematic filing recovers the bulk of them, against a modest $12K/yr of warranty-eligible failures	$8,000
utility rebates claimed	motor rebates run $40–$7,000 per qualifying swap; counting 3 small-motor swaps	$1,500
compliance fines avoided	one lapsed cert or missed renewal	not counted

~$60K in year one: roughly $50K saved, $10K in new cash — against a marginal run cost of pennies of open-model inference per goal plus one browser ticket per work order.

The sensitivity only runs one way. That downtime line is 4 hours at the floor of the published range; the same 4 hours at the top of the discrete band ($50K/hr) is $200K, and every additional hour the overnight queue removes is another $10–50K on top. The recovery lines have the opposite shape: small, certain, cumulative. They are what pays for the tool in cash even if you credit it with zero downtime hours.

The evidence — measured live tonight

Browser agents don't fail because they can't act. They fail because the human didn't say precisely enough what they wanted. We went in expecting OTPs, CAPTCHAs, and login walls, and measured something else. On a live federal form portal, the calibration ticket blocked twice, both times on ambiguity, never once on credentials. Each block burns a full 15–20 minute execution cycle. Re-fired with a fully specified imperative goal, the same task completed with real output (tkt_LclPziYpSgddl0HA-tF3nQ, WIN.md).

gantt
    title the same task, fired twice — the only difference is specification
    dateFormat HH:mm
    axisFormat %H:%M
    section vague ask
    executes, blocks on ambiguity      :crit, v1, 00:00, 18m
    waits for a human answer           :crit, v2, after v1, 12m
    executes, blocks on ambiguity again:crit, v3, after v2, 15m
    fails — no answer                  :milestone, after v3
    section specified ask
    Novita interview + goal compile    :done, s1, 00:00, 1m
    executes straight through          :active, s2, after s1, 17m
    real output                        :milestone, after s2

Loading

The plant run (tkt_os-NZoZVT6Q_-w8vPo7ovA): the specified sourcing goal worked bot-hostile McMaster-Carr for 29 minutes without asking a single clarifying question, then hit the login wall and escalated rather than quitting (blocked_on_user: "provide login credentials…"). That refines the thesis instead of confirming it. Specification eliminates the unbounded ambiguity blocks; what's left on a B2B catalog is auth, which is a bounded block the platform hands to a human and a real facility account resolves. Full read in PLANT.md.

The recovery layer, completed live (tkt_8hDAny1kyyDvj6ewMii77w): the --rebate goal ran pge.com read-only for 67 minutes and came back with the whole filing package — the applicable programs by name (Business Energy Efficiency Rebate; AESAP for clean-water pumps), the dollar amounts ($30 prescriptive flat rate; up to $20/HP ≈ $100 via AESAP), the four documents the application requires, and a deadline it calculated itself: install date 2026-07-10 → file by July 10, 2027. That is money the facility didn't know it was owed, identified overnight from one line of work-order text. No login, no block, no clarifying question — completed end-to-end.

Across suppliers, the split is clean and worth stating precisely. The specificity layer is proven on all of them — McMaster, Grainger, and Zoro each compile a correct, distributor-specific goal in seconds, reproducible offline with --dry and no ticket required. Execution is proven on McMaster to the auth wall described above. A Grainger execution ticket (tkt_fI0oixBm1IWKIkvJKCmHaA) is in flight as of this writing and stays out of the ledger until it's terminal, like every other pending ticket here.

The concurrency envelope (WIN.md): 8 simultaneous tickets all failed in ~3 min, 6 staggered 20 s apart were all cancelled, 1 alone completed repeatedly. That cap doesn't hurt this vertical. A realistic nightly queue of ~20 work orders drains sequentially inside one night shift with the cap exactly as it is. The 15–20 min latency that kills a consumer concierge is irrelevant at 2 AM; the latency selects the use case.

Operating lessons that aren't in anyone's docs:

• Imperative goals succeed, meta-instructions fail. "Complete this form for…" works. "Report how far you got…" dies.
• Instructions truncate around 500 chars, so compile tight goals.
• blocked_on_user detail lives only on GET /tasks/{id}. The documented /v1/actions/tickets/{id} and its MCP tool return nulls.
• Python-urllib's default User-Agent gets 403'd where curl doesn't. Set any UA.
• Novita models are reasoning models: reasoning_content eats the token budget before content exists. max_tokens=400 returns "". Use 3000+.

How it works

Two layers. The fast one asks; the slow one acts.

flowchart LR
    WO["🔧 vague work order<br/><i>&quot;bearing on pump 3 is squealing&quot;</i>"]
    subgraph NOV["specificity layer — Novita · seconds"]
        Q["interview: the 2–4 questions<br/>the senior tech would ask"]
        G["compile ONE imperative goal<br/>&lt;480 chars, uniquely specified"]
        Q --> G
    end
    subgraph AL["execution layer — ActionLayer · 15–20 min"]
        EX["real browser drives the<br/>catalog / portal — doesn't quit"]
        BU["⏸ blocked_on_user:<br/><i>&quot;shielded or sealed?&quot;</i>"]
        EX <--> BU
    end
    WO --> Q
    G --> EX
    EX --> OUT["part № · price · stock<br/>filled cart · filed claim · rebate"]

Loading

The economics write the architecture: never spend a 20-minute browser cycle discovering an ambiguity a 2-second model call could have caught. And blocked_on_user isn't a failure mode, it's the junior tech texting the retired one — at most one question per ticket.

Real output — python3 plant.py "bearing on pump 3 is squealing" --facts workorder.json --dry:

════════════════════════════════════════════════════════════
  THE WORK ORDER
  "bearing on pump 3 is squealing"
  → a catalog search on this blocks or buys the wrong part

  SPECIFIED SOURCING GOAL
  Find a 6203-2RS double rubber-sealed ball bearing (17mm bore, 40mm OD,
  12mm width) on mcmaster.com and return the McMaster-Carr part number,
  unit price in USD, and whether it is in stock—read-only, do not add to
  cart or check out.
  230 chars

════════════════════════════════════════════════════════════
  "bearing is squealing" → wrong part, second truck roll
  exact spec → part number, price, stock — on the bench by first shift
════════════════════════════════════════════════════════════

Real output — plant.py "pump 3 bearing failed, should still be in warranty" --warranty:

  SPECIFIED WARRANTY GOAL
  Go to Grundfos's support site, find the warranty/RMA claim form, and
  complete it with: model CR 3 vertical multistage pump, serial
  GF-2024-118842, purchased 2025-03-14, invoice INV-8841, failure:
  motor-end bearing failure at ~14 months under normal duty; high-pitched
  squeal then seizure. Return the claim reference or the list of fields
  the final submission requires. Do not perform the final submission.
  405 chars · target: https://www.grundfos.com

Real output — plant.py "we swapped pump 3 motor for a premium-efficiency one" --rebate:

  SPECIFIED REBATE GOAL
  Search PG&E's California business rebate pages for programs covering
  replacement of a 5 HP pump motor with a NEMA Premium efficiency motor
  installed on 2026-07-10; return the applicable program name, rebate
  amount, required documentation, and filing deadline. Read-only—do not
  create accounts or submit applications.
  316 chars · target: https://www.pge.com

That 230-character sourcing goal is the retiring tech's translation, done by a model in seconds. Drop --dry and the goal fires as a live browser ticket, with the CLI tailing it to terminal state.

Prove it

python3 verify.py

Re-fetches every ticket cited in this repo from the live ActionLayer API and checks the recorded state still matches. Successes, failures, and cancellations alike — nothing here is asserted from memory. Exit 0 means every claim holds. The ledger is evidence/tickets.json.

  all 12 claims verified against the live API.

Tickets still pending at submission time stay out of the ledger. Check one yourself: ./al.sh get tkt_os-NZoZVT6Q_-w8vPo7ovA

Run it

python3 plant.py "bearing on pump 3 is squealing"      # source the part (read-only)
python3 plant.py "..." --supplier grainger             # same work order, different catalog
python3 plant.py "..." --cart                          # fill cart, stop before the order
python3 plant.py "pump 3 bearing failed in warranty" --warranty   # complete the RMA form
python3 plant.py "swapped pump 3 motor for premium-eff" --rebate  # find the rebate owed
python3 plant.py "..." --dry                           # specificity layer only, ~4s, no ticket
python3 verify.py                                      # re-verify every claim, live

Zero dependencies, Python stdlib only. Needs .env with NOVITA_API_KEY and ACTIONLAYER_API_KEY.

Scope

• Sourcing is read-only. The ticket returns part number, price, and stock. Purchasing is a human click on a filled cart, capped by max_budget_usd once checkout is wired. No checkout evidence yet.
• Symptom→spec is prompted, not yet evaluated. A pilot needs a facility's historical work orders with what was actually ordered as ground truth.
• The dollar model is a model. Every input is at the low end of a published range, but no line has been measured at a real facility. The pilot replaces it with a stopwatch.
• Fortune-500 downtime figures get extrapolated down to small facilities.

The one-sentence pitch

Agents don't need a better browser. They need the two questions an expert would have asked first, so the slow, expensive executor only ever runs a goal that can't block.

The layer where a slow, sequential, never-quits browser agent shines is the paperwork nobody staffs: parts sourcing, warranty claims, rebates, compliance renewals. Value locked behind a form nobody has time to fill. thirdshift is the night clerk that files them.

flowchart TD
    A["🏭 plant work order<br/><i>&quot;pump 3 bearing squealing&quot;</i>"] --> S
    S["specificity layer<br/>the expert's questions, in seconds"] --> E["execution layer<br/>the browser that doesn't quit"]
    E --> R1["sourced part · filled cart"]
    E --> R2["warranty claim · rebate ·<br/>compliance renewal"]
    R1 --> V["💰 value locked behind a form,<br/>filed while nobody's awake"]
    R2 --> V

Loading