Most agent projects work in a demo and break in production. LARS is the set of architecture choices that keeps them working: narrow agents, durable pipelines, gates around everything, and a record of every step.
The architecture
Your agents on top. Everything around them, from LARS.
Lifecycle of a run
What happens between a trigger and a write.
triggerA schedule or the Agent API starts a run and mints one run ID. Overlapping schedules skip or queue by policy.
collectCollector steps call read-only tools through the tool gateway with a per-step token. Results are stored as immutable snapshots and pinned to the run.
evaluateRules run as deterministic code over the pinned snapshots and emit candidate findings with a dedup key. Same input, same findings.
reasonThe analyst agent receives only the candidate and its evidence, with personal data masked. It returns narration, a severity within the rule's bounds and a proposal that must match the schema.
proposeThe proposal is versioned and hashed, routed to the approver role for that project, and the workflow waits. No worker is held while it waits.
decideThe approver signs in with their own identity. The decision is written to the database first, then signalled to the workflow. Timeouts escalate, then expire.
actThe action agent for that action type validates against the tool schema and cost rules, then calls its single write tool. In shadow mode the target is a sandbox.
recordEvery span, model call, tool call, decision and cost record carries the run ID. Audit records cannot be updated or deleted.
Engineering choices
Nine decisions that make the difference.
Narrow agents with a typed output
Each agent has one job, one identity and one output contract. If the output doesn't match the contract, the step fails instead of passing a guess along.
WhySmall agents are testable and replaceable. A mistake can't spread past one step.
Rules decide, AI explains
Where a decision can be a rule, it is code. The model gets the evidence and writes the explanation, a severity within fixed bounds, and a proposal draft.
WhyThe same data always gives the same finding. AI adds understanding, not randomness.
Durable pipelines
Every pipeline runs on a workflow engine that keeps its state. A run can wait three days for an approver, survive a restart or a deployment, and continue exactly where it stopped.
WhyReal business processes take days and involve people. Most agent code assumes seconds.
Approval is a step, not a button
A proposal is versioned and fingerprinted. The approver signs in with their own identity, the exact content they saw is stored, and only that version can be carried out. Timeouts escalate, then expire. They never approve.
WhyWhen an auditor asks who approved what, you have the answer, down to the words on the screen.
Gates are the only way out
Agents never call a system, a model or an email server directly. A tool gateway enforces per-agent allowlists and rate limits. A model gateway masks personal data, caps spend and holds the keys.
WhyPermissions live outside the agent, where no prompt can talk them away.
One run, one ID
Every step, model call, tool call, approval and cost record carries the same run ID. Open any decision and see what was read, what the AI wrote, who approved it and what it cost.
WhyDebugging, cost control and compliance all start from the same trace.
Quality tests before every change
Each pipeline has a set of real test cases that runs on every change to a skill, a prompt or a model. A drop in quality blocks the release. New pipelines start in shadow mode next to the people who do the work today.
WhyEvery new model should make your system better, not break it.
Runs in your cloud, on two machines
Containers on two well-sized servers, one for the runtime and one for observability, with state in a database and object storage. Kubernetes only when you need it.
WhyYour data stays in your account and your region, at a running cost that fits a business process, not a platform team.
A specification, not a template
LARS ships as versioned packages plus a written specification and an automatic conformance check. Updates arrive as normal dependency updates; the check proves your project still meets the spec.
WhyTemplates go stale the day they are copied. A spec and a check stay current.
Open source underneath. Owned by you.
Every component LARS uses is open source, and no feature depends on an enterprise licence. Your code is yours from day one, and the framework can be bought out at a price fixed in the contract.
8 collectors, 3 analysts, a reporter, 5 action agents
The challenge
The health of each client project was spread over hours, invoices, code, cloud costs and deals in different tools. Budget overruns, unbilled work and overloaded servers surfaced late.
What we built
Agents read every system on a schedule and link the data per project. Fixed rules flag budget burn, overdue invoices, hot servers and broken builds. AI explains each finding and drafts a fix: a task, a payment reminder, a server resize. The right person for that project approves it from the dashboard or their phone.
Why it matters for youIt is the same system we demo. In a pitch, prospects approve a real proposal on their own phone and open the full trace of what happened, on a masked copy of our own data.
Built by the LARS team
Likifin
An AI agent for debt collection, on premise.
Sector
Finance
Deployment
Inside the client's own environment
40%
less handle time per call
The challenge
Collection calls are repetitive, regulated and sensitive. The client wanted agents to take on the routine work, without customer data leaving their own environment.
What we built
An agent system that supports collection calls end to end, running on premise. Agents prepare and handle the routine parts; people stay in charge of the decisions that carry risk.
What went into LARSDeployment in the client's own environment is the default, not an upgrade. Personal data is masked before it reaches a model.
Built by the LARS team
BYBD
An outreach engine the client owns, for an industrial recruiter in Rotterdam.
Sector
Recruitment, process industry
Scope
Sourcing, enrichment, scoring, outreach via LinkedIn and email, reply handling
The challenge
BYBD wanted its own outreach engine instead of renting a stack of tools, with data and logic it controls.
What we built
A pipeline that finds companies likely to hire, enriches and scores them, and runs personal outreach sequences. AI is deliberately held to two jobs: writing the one-to-one message and reading hiring intent. Sourcing, scoring and routing stay plain, predictable code.
What went into LARSNarrow agents, and rules wherever a rule will do. AI where it adds value, code everywhere else.
Your process could be the next case.
Start with a free half-day Agent Discovery and leave with a designed pipeline and a fixed pilot quote.
Every LARS engagement is priced on the same three lines, so you always see what you pay for and why. We quote after a Discovery, once we know your process.
Built for you
Project
Your agents, skills, pipelines, connections and dashboard. Estimated per pipeline and per tool, delivered as a pilot first, then production.
LARS framework
Licence
A one-time setup and a recurring licence for the framework and its updates. Optional operations if you want us to run it for you.
External services
Pass-through
AI model usage, hosting and third-party services, at cost plus a management fee, with budget caps on model spend.
Buy out any time
Your code is yours from day one. The framework can be bought out at a price fixed in the contract, lower with every month of licence paid. You receive a perpetual licence with the full source.
Start small
A half-day Discovery, then a pilot on one pipeline with real cases. You only move to production when the pilot proves its value.
Talk to sales
Book a free Agent Discovery.
Half a day with your team, on site or online. You leave with your top three processes for agents, one pipeline designed, a business case and a fixed pilot quote. Free until the 2027 budgets close.
Email[sales@lizard.global]
Phone+60 10 213 4037
OfficesRotterdam · Kuala Lumpur · Bali
Part of Lizard Global, building software since 2013.
Common questions
Why don't you publish prices?
The project line depends on your process and systems, and we only quote what we have scoped. The Discovery gives you a fixed pilot quote within a week.
Do we need our own team to run LARS?
No. We can operate it for you, or hand it over to your team. Both are part of the model from the start.
Which AI models can we use?
The providers you approve, switched by configuration, or models hosted in your own cloud. Model spend is capped per pipeline.
What happens if we stop working with Lizard Global?
Your code is already yours. Buy out the framework at the contracted price and keep running everything with your own team, without any dependency on us.