AI Agents & Autonomous Systems for Operational Scale

Autonomous systems
for operational scale.
Descend
THESIS
The interesting work is no longer at the surface. It's underneath.
Most of what's sold as “AI” is a chat box wrapped around someone else's API.
We build the layer below that — the orchestration, the state, the recovery paths, the operational logic. The system that, when it runs, no one notices.
DISCIPLINE
Production engineering applied to autonomous behavior. No demo-ware.
STANCE
Systems should be observable, reversible, and operated by humans who understand them.
CADENCE
Two systems concurrent. We do not bid on more.
OUTCOME
An infrastructure that survives leadership changes, vendor changes, and the next model release.
Read our field notes →
OPERATIONAL LAYERS
Six layers of autonomous operational depth.
Every Gruntflow engagement is composed from the same primitive layers — ordered by depth, not by category. We don't ship features. We ship operational substrate.
LAYER · 01ACTIVE

Signal Acquisition

Continuous monitoring of public surfaces, conversation streams, and intent signals. Behavioral throttling and human-shaped pacing are first-class.

94 sources · 14ms median→
LAYER · 02ACTIVE

Multi-Agent Coordination

Specialist agents that plan, critique, and hand off. Memory, retrieval, and tool-use under a planner that owns the operational graph.

Planner / Critic / 12 tools→
LAYER · 03ACTIVE

Workflow Intelligence

State machines that enforce lifecycles, audit transitions, and gate quality. Built around the operational semantics — not generic ticket queues.

State / Lifecycle / Audit→
LAYER · 04ACTIVE

Voice Interfaces

Sub-second voice loops with conversational memory and live interaction states. Designed for noisy real-world rooms — not demo conditions.

< 480ms turn latency→
LAYER · 05ACTIVE

Human-in-the-Loop

Operators stay in command. Approval surfaces, escalations, override paths, and reversible automations are part of the protocol — never bolted on.

Reversible by design→
LAYER · 06ACTIVE

Operational Infrastructure

AWS-grade backends, observability, role-based logic, and data flows engineered for the long run. Boring, durable, instrumented — the way real systems live.

AWS / observable / SOC-ready→
SYSTEM DOSSIERS
Four systems, in production.
Each runs continuously. Each is owned by an operator who can read every transition. Each was built to a single specification: this should still be running in three years, after the team that commissioned it has changed.
// 001PROJECT NORTHWIND

Continuous warm-lead acquisition across public surfaces.

A signal-acquisition system that watches public conversation for buying intent, qualifies prospects, and performs human-shaped outreach with adaptive throttling and behavioral simulation. Integrates as an additional sales rep — not a spam cannon.
Domain
Lead Acquisition
Stack
Python · Meta APIs · GPT-4o
Latency
14ms detect / 90s respond
Status
IN PRODUCTION
NORTHWIND · LIVE FEED◉ SIG-ACQ
AOI: 4 / SIGNALS: 124kWARM: 38
// 002PROJECT MERIDIAN

A multi-agent content engine that doesn't break.

Custom marketing automation infrastructure: a planner agent coordinates ingest, brief, write, render, critique, and publish specialists — under analytics feedback that retunes the brief. Cloud-native backend, video generation pipelines, Meta API integrations.
Domain
Multi-Agent Orchestration
Agents
6 specialists / 1 planner
Throughput
~140 assets / week
Status
IN PRODUCTION
MERIDIAN · ORCHESTRATOR◉ MAS
PIPELINE: STABLEQUEUE: 12
// 003PROJECT ATRIUM

Voice that answers, listens, and remembers.

A real-time conversational AI platform. Sub-second turn-taking, conversational memory across calls, multi-turn context, and adaptive responses. Built for production rooms — variable network, accents, interruptions.
Domain
Voice Interface
Turn Latency
< 480ms median
Memory
Cross-session, scoped
Status
IN PRODUCTION
ATRIUM · CALL 47-A◉ VOICE
TURN 04 · CTX 2.1kRTT 412ms
// 004PROJECT BASELINE

An operational backbone for a 600-tech field service.

A state-driven service platform: lifecycle enforcement, technician coordination, audit trails, revisit tracking, role-based operational logic, and analytics pipelines on AWS. Built around the actual semantics of HVAC service, not generic ticketing.
Domain
Operational Platform
States
8 enforced lifecycles
Scale
600 techs · 14 regions
Status
IN PRODUCTION
BASELINE · LIFECYCLE◉ STATE
JOBS LIVE: 312SLA 99.6%
FIELD DISPATCH
Notes from systems in production.
We publish a short observation from inside each engagement, on a slow cadence, with as little editorialising as the work allows. Audit-log honesty over case-study polish.
— ENTRY · 001 —

A field tech told us today that the system has changed how she thinks about a callback.

We didn't design for that. We designed for a sequence of state transitions that close the lifecycle of a service call cleanly. What she described was something else — an internalised expectation that the loop would close, which freed her attention to do the actual repair. Tools shape the people who use them. We are watching what this one propagates, with caution, in writing, on a slow cadence.

— ENTRY · 002 —

The caller had a question we had not anticipated, asked in a way the model had not encountered.

It answered well, then asked a follow-up that surfaced a constraint the team had been arguing about internally for two days. We replayed the turn six times. It was not the model being clever; it was the conversational graph routing through a node we had built and forgotten. Most of what looks like intelligence, in production, is the architecture working.

— ENTRY · 003 —

We spent the week building the system that refuses to act.

The detection layer is now tuned to a confidence band wide enough that a human reviews any signal that lands on its edge. Throughput dropped by 41%. Conversion quality moved up by a larger margin than that. Restraint is a feature you engineer. The platform doesn't reward it, but the operator does, and the operator is who we work for.

— ENTRY · 004 —

The critic agent rejected the asset our human team had approved.

We pulled the trace. The critic had caught a tonal mismatch with a brand guideline added six weeks earlier — a guideline the human reviewer had not yet internalised. We do not present this as a triumph. The critic is well-prompted because we wrote it; it found the seam because we logged the guideline. The agent gets the credit; the audit trail gets the operator home.

ARCHIVE OPEN ON REQUESTEST. 2021 · 47 ENTRIES FILED
ENGAGEMENT — INTAKE

We onboard two systems per quarter.
The intake is a conversation, not a form.

Send a brief on the operational system you'd like built. We respond in 48 hours with a written read of the problem, an architecture sketch, and a yes or a referral. We don't pitch.