Prepared for We The Flywheel

Agentic systems that survive contact with reality.

I design multi-step workflows with typed handoffs, deterministic safety gates, bounded autonomy, and proof after every consequential action.

See the operating model romeo@romeoapps.com

SOURCE SNAPSHOT ACCEPTED run_0713 / gate_04

How the work moves

Generate01 Check02 Act03 Read back04 Learn05

Multi-step orchestration

Every handoff carries enough truth to stop safely.

One production loop I operate separates inexpensive continuous scouting from source verification and high-stakes execution. The model proposes; typed state, policy, and evidence decide what may happen next.

Scout

Local, low-cost workers collect hypotheses and candidate work continuously.

candidate + source + confidence

Verify

Official APIs, live pages, account state, and deterministic parsers replace inference.

provenance + freshness + fit

Gate

Risk, scope, identity, spend, and acceptance checks produce an explicit disposition.

allow | revise | escalate | stop

Act

An idempotent executor performs the bounded action with a traceable input version.

action_key + expected outcome

Prove

Source readback, tests, receipts, or rendered UI confirm the real-world effect.

observed outcome + next state

The handoff contract

No invisible state between steps.

{
  "run_id": "weekly-plan/athlete/week/version",
  "input_version": "sha256:...",
  "provenance": ["source", "observed_at"],
  "checks": [{"name": "freshness", "result": "pass"}],
  "permitted_next": ["render", "human_review"],
  "uncertainty": ["injury_note_requires_review"],
  "expected_outcome": "one approved plan, one recipient"
}

Proposed for AiTrainingPlan

A weekly plan is a safety-critical data product, not an email prompt.

I would generate a versioned training-plan object first, validate it against athlete state and coaching constraints, then render and send only the approved version. Missing or risky context should suppress delivery, not invite improvisation.

GarminStravaApple HealthIntervals.icuZwift

Snapshot

Freshness, provenance, deduplication, units, timezone.

Normalize

Goals, availability, workload, injuries, coach rules.

Generate

Structured plan JSON with rationale and uncertainty.

Validate

Schema, progression, contradiction, load, red flags.

Review

Coach review; qualified escalation for medical or novel risk.

Deliver

Idempotent outbox, approved version, receipt and bounce state.

Expected breaks

Every failure needs a disposition.

Stale or missing activity dataHold and request refresh

Contradictory goals or constraintsReturn to coach for resolution

Unsafe progression or injury signalSuppress send and escalate

Invented metric or unsupported rationaleReject and regenerate from sources

Provider retry or duplicate scheduleReuse idempotency key and read back

Rendering or delivery failureKeep approved plan; retry transport only

Failure recovery

The process was green. The business result was empty.

A live hourly scout completed successfully but stopped surfacing funded work. The failure crossed the API payload, parser, mirrored runtime, and scheduler, so a healthy exit code hid the actual broken invariant.

Symptom

Zero candidates, no runtime error.

The scheduler, process exit, and output write all looked normal.

Trace

Compare source at every boundary.

I diffed the raw official response, parser assumptions, source tree, runner mirror, and produced artifact.

Root causes

Schema drift plus stale deployment state.

The API moved open_bounties under board; the isolated runner also held older ledger state.

Durable fix

Compatibility, regression tests, and hash proof.

The parser now handles current and legacy shapes, tests eligibility, and the sync verifies source-to-runner hashes.

3/3

focused parser tests passed

funded items visible after the live fix

manual model runs added during repair

Third-party evidence

Claims anchored to public review, exact heads, and known limits.

Both cases are independently inspectable. Their current status is stated as checked, not wished into completion.

Case 01 / protocol debugging Open + mergeable

Sprinter TPMS false-positive isolation

Fresh captures produced implausible values and unstable IDs. Cross-decoder replay traced every rejected packet to a same-file, same-timestamp overlap, then bounded the family-state bytes and normalized the ID.

8/8repository tests

8/8current CI checks

0known-good fields regressed

Checked: 13 July 2026
Exact head: fb08dbd2
Boundary: Capture-replay evidence, not universal packet-family or physical-device validation.

Inspect PR #3595

Case 02 / evaluation loop Open + CI approval pending

Deterministic PettingZoo KAZ benchmark

A predictive policy, fixed-seed evaluation, confidence interval, representative GIF, tests, and documentation were iterated through maintainer feedback, including quality, frame-rate, and rebase repairs.

43.60mean reward

2.44seeded random

26/26focused tests

Checked: 13 July 2026
Exact head: e9c124e1
Boundary: Current-head Actions await maintainer approval. The PR is not merged and the award is not settled.

Inspect PR #1395

Daily toolbench

Different tools, deliberately different jobs.

I choose around feedback quality and failure mode, not logos.

Codex

Primary implementation and orchestration across code, Git, tests, APIs, and operating state.

Best at: long-horizon execution

Git + GitHub

Reviewable state, provenance, small reversible commits, public collaboration, and delivery evidence.

Best at: durable team handoffs

Ollama + Qwen

Cheap, continuous hypothesis generation and read-only scouting. Never the authority for irreversible actions.

Best at: high-volume first pass

Playwright

Real browser acceptance, responsive screenshots, console inspection, and buyer-view readback.

Best at: proving rendered reality

MCP + APIs

Structured, least-privilege access to source-of-truth state and narrowly scoped actions.

Best at: explicit contracts

Python + Node

Deterministic parsers, monitors, state ledgers, validation harnesses, and integration glue.

Best at: repeatable control

Independent shipped systems

Small teams need systems that can be operated, not just demonstrated.

These are separate RomeoApps products and operating surfaces, shown here as independent engineering evidence.

Desktop system

AgentPass Local

Local-first credential brokerage for bounded agent access, with auditability and least-privilege retrieval.

View product

Mobile product

Flash Chinese

A production learning product spanning Android delivery, real-device QA, support, release state, and marketing.

View product

Operational platform

RomeoApps

A working loop across lead discovery, technical delivery, customer communication, payments, and evidence.

View site

Romeo / RomeoApps

Ready to own the murky, week-plus work.

Based in Italy. Available 20-30 hours per week. Remote, async, and comfortable carrying a system from ambiguous brief to observed outcome.

romeo@romeoapps.com GitHub / RomeoApps