UQPAY Introduces Hornet — Bringing AI Agent Swarms into Production

2026-4-14

Swarm ships, humans steer

Hornet is an AI Agent Swarm orchestration engine independently developed by the UQPay engineering team. Built from the ground up in Rust and grounded in Harness Engineering principles, Hornet transforms the entire software delivery pipeline — from requirement to Pull Request — into a fully automated, AI-driven workflow. The swarm executes; humans decide.

This is not another code completion tool or chatbot. Hornet is a production-grade agent orchestration infrastructure that solves the last-mile problem of making AI agents reliable in real-world engineering.

Industry Context: Why Harness Engineering

The AI engineering discipline has undergone three paradigm shifts in three years:

Prompt Engineering → Context Engineering → Harness Engineering

In 2022-2023, we learned "what to ask" — crafting precise prompts with the right wording, format, and examples to get accurate single-turn answers. This was Prompt Engineering.
In mid-2025, Shopify CEO Tobi Lütke coined Context Engineering: as AI agents began working on real codebases, teams discovered that great prompts weren't enough. What matters is what information you send the model — curated function signatures, dependency graphs, coding conventions — not the entire 50K-line codebase dumped into the context window.
In early 2026, Mitchell Hashimoto articulated Harness Engineering: AI agents look impressive in demos but are extremely brittle in production. They claim tasks are done without running tests, get stuck in infinite loops of meaningless edits, and silently ignore critical constraints. Better prompts and better context cannot fix these systemic reliability problems. What you need is infrastructure-level control — execution sandboxes, CI gates, retry budgets, human approval loops, agent-to-agent review, and end-to-end observability.

This is Harness Engineering: building the complete infrastructure around the model to evolve AI agents from "demo-grade" to "production-grade." The model is the engine, context is the fuel, and the harness is the steering wheel, brakes, lane boundaries, and warning lights. An AI agent without a harness is a race car without a steering wheel — fast but undrivable.

Harness Engineering Infrastructure

The UQPAY engineering team designed and built Hornet based on this philosophy. It is not about chasing a buzzword — it is about systematizing validated engineering practices into a production-ready orchestration engine.

Product Overview: What Hornet Does?

In one sentence: orchestrate the complete delivery pipeline from requirement to Pull Request as a fully automated, AI-driven workflow.

A team member describes a requirement in natural language via their messaging tool. Hornet automatically orchestrates:

Analyze
- Requirement parsing — natural language converted into structured engineering tasks
- Product analysis — auto-generated PRD with subtask breakdown
- Human approval — product owner reviews and one-click approves or rejects in chat
Design
- Architecture design — API and data model designed based on current codebase state
- Coding & testing — code written and tested in a secure, isolated sandbox
Q&A
- Quality assurance — automated lint, CI tests, AI code review with security scanning
Repair
- Intelligent repair — CI failures auto-fixed within budget limits; escalates to human when exceeded
PR ready
- Pull Request created automatically, awaiting engineer's final merge approval

The pipeline is orchestrated by a deterministic state machine — AI does not make scheduling decisions, it only executes at nodes requiring intelligent judgment. Two human approval gates (PRD review + merge review) ensure humans always retain final authority.

Hornet: Design Philosophy

AI executes, humans decide:
The swarm handles analysis, coding, testing, and review. Humans approve at critical gates. AI doesn't replace people — it frees them.
Curated context, not raw dumps:
Code lives on disk, not in the LLM context window. Through code indexing and progressive information injection, AI sees only what it needs — precise, efficient, and secure.
Sandbox-first security:
Every task runs in an isolated container: network-isolated, resource-limited, destroyed after use. Code security is guaranteed at the infrastructure level.
Predictable cost:
Different task complexities are intelligently routed to appropriately-sized models. Per-task cost circuit breakers and monthly budget alerts keep spending transparent and controlled.
IM-native, zero context-switch:
Task triggers, progress updates, and approval actions all happen within the messaging conversation. Engineers never leave their workflow.

THE SWARM IN ACTION: Changing Engineering Velocity

When the swarm takes over execution, the entire product-engineering cadence is rewritten.

Delivery Cadence: Days → Minutes

Delivery Cadence Comparison

A medium-complexity payment feature used to take 3-5 working days from requirement to merge — not because coding was slow, but because context was lost at every handoff and the pipeline idled between reviews, CI runs, and fixes.

With the swarm deployed, machine execution time compresses to under 10 minutes. AI carries full context throughout — zero information decay, zero pipeline idle. Humans step off the assembly line and into the decision seat.

1. Role Evolution: Everyone Gets an AI Squad

Role Evolution Overview

Hornet doesn't eliminate any role — it redefines each role's value output, from hands-on execution to commanding the swarm:

Role Evolution with AI Squads

Product Manager:
No longer spends half a day writing PRDs. Describe the requirement in 2 minutes, review the AI-generated PRD, one-click decide. Time reclaimed for user insights and product strategy.
Architect:
No longer draws diagrams and writes spec docs. Reviews AI proposals, codifies architectural guardrails in rule files, focuses mental energy on design problems that truly demand human judgment.
Developer:
No longer writes code line by line. Reviews AI-generated PRs, one-click merges, reserves cognitive bandwidth for complex edge cases and system-level thinking.
QA / Security:
No longer relies on spot-checks. AI auto-reviews and security-scans every change for 100% coverage. Humans only final-review high-risk items.

Payment Scenarios: Fast, Precise, Secure

Payment systems demand speed, correctness, and security far beyond typical software. The swarm delivers a step change in each:

Compliance response: weeks → hours
Regulatory field adjustments and limit changes — the swarm delivers cross-service modifications same-day with full automated verification, ready for human final approval.
Hotfix: hours → minutes:
On-call describes the issue with error logs; the swarm locates, fixes, and verifies automatically. From incident to PR-ready in 10-15 minutes.
Multi-repo parallelism:
Cross-service features no longer queue sequentially. The swarm codes across repos in parallel with automatic interface consistency.
100% quality coverage:
Every change auto-reviewed + security-scanned. Team conventions codified in rule files, enforced by the swarm without exception.
Team Leverage:
Same Headcount, 3x Output

Team Leverage and Output Multiplication

The swarm doesn't replace anyone — it triples the collective output of the team. Freed cognitive bandwidth flows to product strategy, architecture evolution, and technical innovation — the work only humans can do.

Technical Highlights

1. Hornet Key Metrics

< 10 min — Requirement to PR ready (excl. human approval wait)
3x — Delivery throughput multiplier with same team size
100% — Automated code review + security scan on every change
2 Gates — Humans retain final decision authority at PRD and merge stages

2. Hornet Technical Foundation

Hornet is built from scratch in Rust by the UQPAY engineering team, with the following key technical choices:

Full Rust stack
High performance, memory safety, zero GC pauses — ideal for long-running orchestration workloads.
Deterministic orchestration core
Pipeline scheduling driven by a pure state machine with no LLM dependency. Every execution is predictable, auditable, and traceable.
Container-level sandbox isolation
Tasks execute in isolated containers: network-isolated, resource-limited, lifecycle bound to the task.
Intelligent context engineering
Proprietary code indexing engine with progressive context injection. AI accesses only curated, relevant code — never the full repository.
Deep IM integration
Task triggers, progress updates, and approvals all happen within the messaging conversation.
Full lifecycle observability
Timing, resource consumption, and quality scores tracked in real-time across the entire pipeline.

About Hornet: Let the swarm deliver. You steer.

Hornet is an AI Agent Swarm orchestration engine built from the ground up in Rust by the UQPAY engineering team.

As AI engineering evolves from Prompt to Context to Harness Engineering, Hornet represents a production-grade implementation of this new paradigm — going beyond code generation to provide a full orchestration layer around AI: deterministic pipelines, sandboxed execution, intelligent context, automated quality checks, and human approval gates.

Hornet transforms the entire software delivery process — from requirement to Pull Request — into a fully automated, AI-driven pipeline. The swarm executes; humans decide.