Control and orchestration
Define agent state, tool routing, approval points, retries, cancellation and the boundary between deterministic code and model decisions.
Capability · AI agent systems
Engineering the system around the model
Neul Labs designs the control, state, tool, data and evaluation layers that turn a model call into a bounded software system. The goal is not maximum autonomy. It is useful behaviour with observable decisions, recoverable failures and clear human authority.
What the work covers
An engagement starts with a specific user, decision and failure budget. We map the current workflow before choosing a model, framework or orchestration pattern.
Tool and data boundaries
Expose the smallest useful capability, validate every argument, apply tenant and field policy, redact secrets and record attributable actions.
State and recovery
Design checkpoints, idempotency, durable events and recovery paths so a failed tool call or worker does not silently corrupt the workflow.
Evaluation and operations
Turn expected behaviour into repeatable tests, trace the important transitions and set explicit latency, cost, security and human-review gates.
Engagement contract
Inputs, outputs and acceptance
A useful agent project makes the operating boundary testable before implementation expands.
- Inputs
- Target users, example tasks, prohibited actions, systems and data involved, expected volume, current failures, latency and cost constraints.
- Design record
- State model, tool contracts, identity and permission boundaries, approval points, retention, recovery and ownership decisions.
- Implementation
- A narrow reference path or integration with tests, traces, safe configuration, rollback and documented assumptions.
- Acceptance
- Agreed success and failure scenarios, security and privacy checks, measured latency and cost, runbook and human sign-off.
Inspectable evidence
Related public engineering work
These repositories expose relevant design choices for review. They are examples of Neul-owned engineering work, not customer case studies, warranties or proof that a component fits every deployment.
Rust
brat
run several coding agents in parallel, safely.
Rust
agentvfs
sandbox an agent’s files and shell.
Python
ormai
let an agent query a database, safely.
Python
fastagentic
ship a local agent to production as an API.
Python
memorg
add durable memory to an LLM chat app.
Rust
mcp-pay
charge for MCP tools across payment rails.
A good fit
- There is a named workflow, owner and measurable failure or opportunity.
- The agent needs controlled access to tools, data or long-running state.
- A prototype exists but needs evaluation, recovery, observability or hardening.
- The team wants an evidence-backed build-versus-buy or framework decision.
Not a good fit
- The goal is an unsupervised replacement for accountable human decisions.
- No one can define allowed actions, data authority or a useful evaluation set.
- The requirement is only a generic chatbot or a predetermined vendor purchase.
- Success depends on guarantees about model accuracy that no test can support.
Questions people ask
What is AI agent systems engineering?
AI agent systems engineering is the design and implementation of the software around a model: state, tools, permissions, memory, human approvals, evaluation, observability, failure recovery and operating controls. It is broader than writing prompts and narrower than claiming that an agent can autonomously replace an organisation.
Can Neul Labs work with an existing agent framework?
Yes. A project can keep its current framework when the framework is not the main constraint. The first step is to map the actual workflow and failure modes, then decide whether to extend, isolate, accelerate or replace individual components.
How do you evaluate an AI agent?
We define representative tasks and failure cases, use synthetic or authorised data, pin versions, record tool calls and outcomes, measure task success, invalid actions, recovery, latency, token and compute use, and report limitations. Model-based judges can supplement but do not replace deterministic checks and human review.
Does a public Neul project prove production readiness?
No. A public repository demonstrates inspectable engineering work. Production readiness depends on the selected version, workload, deployment, data, security, support and acceptance tests for the specific engagement.
Start with one workflow and its hardest failure
Share the current system, intended user, tools, data boundary and what must never happen.
admin@neullabs.com