WATCHD
A MARKET FOR MACHINE LABOR
FIG. 01 / THE THESIS

Fund the loops that prove they help. 

Agents are markdown files with a price. A deterministic allocator decides which one deserves the next dollar. You rate what helped. The budget learns.

ONE GO BINARYPLAIN FILESNO MODEL JUDGES THE PORTFOLIOMIT
THE ALLOCATIONMARKET OPEN
00:30✓uptime $0.008all healthy, 200 OK in 142ms01:00★outcome usefulexposed a pricing move, allocation raised02:00✓security $0.214raw SQL in /search, patch attached02:41→competitor skippedagent has unrated output04:15✓repo-health $0.21verify passed, auto-rated useful05:12⊘dbcleanup heldplan pending your approval06:00✓digest $0.089morning brief compiled from 5 agents06:01▲portfoliospend shifted toward verified return
FIG. 02 / THE DOCTRINEFIVE CLAUSES
§ 01

Cron believes every job is equally valuable, forever. 

watchd does not. verified outcomes per dollar decide who runs.
§ 02

A schedule is eligibility, not entitlement. 

due means allowed to compete. the allocator decides what runs.
§ 03

Useful compounds. Harmful starves. 

ratings are append-only. corrections are the compounding asset.
§ 04

No model judges the portfolio. 

one inspectable formula. identical state, identical decision.
§ 05

Anything dangerous waits for you. 

observe → propose → act. authority is a ceiling, never a default.
FIG. 03 / ONE NIGHTSCROLL TO ADVANCE
LOCAL TIME
23:00

While you sleep, the allocator decides which due agent deserves the next dollar.

You slept. They didn't.
6 RUNS · $0.37 TOTAL · 1 PLAN AWAITING YOUR APPROVAL
23:08» watchd up
5 agents armed. schedules are eligibility, not entitlement.
00:30✓ uptime$0.008
All healthy. 200 OK in 142ms.
01:00✓ errors$0.021
3 new timeouts in the payment worker. Same root cause.
02:00✓ security$0.214
New commit adds raw SQL in /search. Flagged with a patch suggestion.
04:15✓ competitor$0.041
Acme cut Pro pricing 20%. Second cut this quarter. Reading as a price war.
06:00✓ digest$0.089
Morning brief compiled. 4 findings, 1 plan pending your approval.
FIG. 04 / THE ALLOCATORDETERMINISTIC

One formula decides. You can read it. 

Every scheduled agent is a strategy competing for a finite daily budget. A bandit, not a vibe: no model scores the portfolio, no prompt decides who runs. Four terms, all inspectable.

score=weight×(expected+ε·uncertainty)/cost
internal/portfolio/portfolio.go · HOVER THE TERMS
WEIGHT
goal importance
Declared in the goal file. product: 3, chores: 1. Money follows what you said matters.
EXPECTED
(1 + useful − harmful) / (2 + rated)
Laplace-smoothed verified return. New agents start humble at one half; every rating moves the posterior.
ε · UNCERTAINTY
√( ln(N + 2) / (n + 1) )
A bounded exploration bonus, UCB-style. Undertested agents keep getting auditions, so an incumbent never owns the budget.
COST
actual average spend
The denominator is reality: observed dollars per run, not the number the agent promised. Floor at one cent.
EVERY DECISION SHIPS A REASON:✓ highest verified return→ agent has unrated output→ pending review cap reached→ global daily budget exhausted→ goal daily budget exhausted→ unrated review cap reached
IDENTICAL STATE PRODUCES THE IDENTICAL DECISION. EVERY ADMISSION REASON IS STORED, QUERYABLE, AND YOURS TO AUDIT.
FIG. 05 / THE LEDGERAPPEND-ONLY

Three words train the market. 

useful, neutral, harmful. Ratings append, never overwrite, so your corrections stay auditable and become the portfolio's compounding asset. Where a shell command can check reality, you don't even have to type them.

RATE THE NEXT RUN OF competitor ↓
expected = (1 + 1 − 0) / (2 + 2)= 0.500
TOMORROW'S ALLOCATION MOVES WITH THIS NUMBER
competitor_0609usefulexposed a pricing move
competitor_0610neutralno change detected
OR LET REALITY DO THE RATING ↓
---
name: repo-health
goal: product
verify: go test ./...
verify_timeout: 2m
---
satisfiedalready true before the run: $0.00, no tokens burned
usefulfalse before, true after: auto-rated, no human needed
incompletestill false after execution: recorded neutral
errorverifier failed or timed out: quarantined from the stats
THE VERIFIER RUNS BEFORE AND AFTER. ALREADY TRUE MEANS NO MODEL IS EVEN WOKEN UP. ITS OUTPUT IS CAPPED AT 8 KiB AND TREATED AS UNTRUSTED DATA.
BACKLOG? watchd rate TRIAGES EVERYTHING UNRATED, ONE KEYPRESS PER RUN. THE ALLOCATOR HOLDS AGENTS WITH UNRATED OUTPUT, SO RATING THROUGHPUT IS THE MARKET'S LIQUIDITY.
FIG. 06 / AUTHORITYobserve / propose / act

Power is granted, never assumed. 

Goals set a ceiling; agents can only lower it. A gated run gets read-only tools and must end with a concrete plan. Nothing executes until you approve, and notify pushes the plan to your phone instead of waiting to be noticed.

observeread-only tools. looking is free.
proposeread-only, plus a plan held for your approval.
actconfigured tools. gate: true still demotes it to propose.
---
name: dbcleanup
schedule: 1d
model: sonnet
gate: true
notify: "ntfy pub alerts 'watchd: $WATCHD_AGENT \
  $WATCHD_STATUS'"
---
 
# DB Cleanup
 
Find bloated tables, unused indexes, and
rows older than the retention policy.
Propose a cleanup plan.
$ watchd pendingAWAITING APPROVAL
ID    AGENT      STATUS
a3f9  dbcleanup  pending (2m ago)
PLAN · READ-ONLY PASS
1. VACUUM ANALYZE on 4 bloated tables
2. Drop unused index idx_sessions_legacy
3. Archive 48,210 rows >180d from events
APPROVING RESUMES THE SAME SESSION, SO IT EXECUTES EXACTLY THE PLAN YOU READ. EVIDENCE IS RECHECKED FIRST, SO A RECOVERED SYSTEM SUPERSEDES A STALE PLAN. TRY IT ↑
FIG. 07 / MEMORYmemory: true

Loops that compound. 

One line of frontmatter gives an agent a notes file it curates itself: injected at the start of every run, rewritten at the end. A scanner that re-reports the same findings is noise. One that remembers builds a position.

RUN 0123:30
memory/competitor.md — memory is empty — the agent writes a baseline
## Baseline — 3 competitors
- Acme: Pro $49/mo, AI addon beta
- Initech: usage-based, no free tier
- Globex: enterprise only, POC req'd
RUN 0205:30
memory/competitor.md — reads the baseline, reports only the delta
## Baseline — 3 competitors
- Acme: Pro $49/mo, AI addon beta
- Initech: usage-based, no free tier
- Globex: enterprise only, POC req'd
+ 06-09: Acme launches annual billing
+ Initech free tier rumored (HN)
RUN 0311:30
memory/competitor.md — connects runs into a trend you can act on
## Baseline — 3 competitors
- Acme: Pro $49 -> $39 (-20%)
- Initech: usage-based, no free tier
- Globex: enterprise only, POC req'd
- 06-09: Acme launches annual billing
+ 2nd Acme price cut this quarter
+ pattern: price war forming.
+ watch for the Globex response
CURATED, NOT TRANSCRIPT-STUFFED. STALE ENTRIES GET PRUNED, AND A POISONED PAGE SCRAPED IN RUN 12 NEVER BECOMES STANDING INSTRUCTIONS FOR RUN 13
FIG. 08 / WHY NOT CRON + BASH

Cron runs scripts. watchd runs judgment. 

A bash script checks if the endpoint returned 200. An agent notices the 200 took four seconds, that the body is an error page wearing a success code, that the same timeout pattern showed up last Tuesday. You write intent in plain language; the model does the interpreting; the allocator pays for what proves out.

CRON + BASH
✗Every recurring job assumed equally valuable, forever.
✗Write a bash script. Parse output yourself. No structure.
✗No cost tracking. You find out at the end of the month.
✗No run history. stdout is gone when the terminal closes.
✗Runs with your full permissions from minute one.
✗No memory. Every run starts from zero and re-reports.
WATCHD
✓Budget flows to verified value. Schedules are only eligibility.
✓Write a markdown file. The model handles the thinking.
✓Cost tracked per run. Budgets enforced mid-run.
✓Every run a queryable record with provenance hashes.
✓observe / propose / act: authority is a ceiling.
✓memory: true means runs build a position and compound.
SAME VERDICT FOR CLOUD ROUTINES: A HANDFUL OF STATELESS RUNS PER DAY, HOUR FLOORS, NO APPROVAL PRIMITIVE. WATCHD LOOPS ARE UNLIMITED, SUB-MINUTE, REMEMBER EVERYTHING, AND WAIT FOR YOU.
FIG. 09 / UNDER THE HOODSINGLE BINARY

2,600 lines of Go. The model thinks. The math chooses. 

No AI runtime, no database, no API keys to manage. watchd spawns claude -p, parses the JSON, and records cost, evidence, allocation, outcomes, and instruction hashes. Every run answers "which instructions produced this output": prompt_hash and agent_hash travel with the record.

cmd/watchd
Entry point. One binary, no runtime deps.
internal/cli
run, up, portfolio, outcome, pending, approve
internal/agent
Parses markdown agents + YAML frontmatter
internal/portfolio
Goals, policy, review limits, deterministic allocation
internal/runner
Spawns claude -p: memory, gate, verify, notify
internal/store
Run history as JSON files with provenance hashes
internal/daemon
Scheduler loop for recurring agents
≈ 2,600 LINES OF GOONE DEPENDENCY: yaml.v3PLAIN FILES, NO DATABASE
cli -> runnercli -> storecli -> agentcli -> daemoncli -> portfoliodaemon -> runnerrunner -> agentrunner -> storeportfolio -> agentportfolio -> store
FIG. 10 / OPEN THE MARKET2 MINUTES

Fund your first loop tonight. 

Star on GitHub
BUILT BY © LEVEL09