Back to Software
AI agent building blocksAvailable

@msm-core SDK — the agent runtime

The execution loop, memory, guards and output — thirteen published packages you wire once and ship.

Memory here is not a chat log: five tiers — customer, episodic, semantic, procedural and reflection — each behind its own storage port that you wire to whatever you already run.

Long missions are a durable state machine: exactly-once execution, cron scheduling, and a human-approval stop before any consequential action. The agent neither works silently nor fails silently.

Running — how the agent works

mini — the execution loop

The heart of the agent: brain calls, guards, tool dispatch, a cost cap it cannot exceed, and a forced finalize when it runs long.

planner — verify and adapt

Breaks the goal into steps and checks each one, retrying or escalating to a human — and it knows which steps have irreversible effects, so it never repeats them.

jobs — durable missions

A state machine with guaranteed exactly-once execution, cron scheduling, and a human-approval stop before any consequential action.

Knowledge — what the agent knows

memory — five tiers

Customer, episodic, semantic, procedural and reflection — each behind a storage port you wire to whatever you already run.

context + rag — context and retrieval

Chunk, embed and search, then assemble tiered context inside a computed token budget — with zero runtime dependencies.

learning — it learns from its runs

Turns run outcomes into reusable lessons, so the agent doesn't repeat the same mistake next time.

Guards — what is not allowed

validate — the output gate

Explicit rules, then embedding similarity, then an LLM judge when needed — and when the gate is unsure it fails to human review, never through.

safety — before anything leaves

PII redaction (card numbers Luhn-checked), prompt-injection detection, and groundedness checks against the source.

gate — the human approval

An explicit approval stop before any consequential action, with auto-approval under conditions you write in plain language rather than code.

Output — what it delivers

parse — file to text

Cloud-free parsers and born-digital PDF handling — your documents are not uploaded to an outside service to be read.

docx — real Word documents

Native Word XML, not renamed HTML — RTL and Arabic aware, carrying your company letterhead.

finance — deterministic maths

NPV, IRR, payback and sensitivity analysis — computed by code, never by the model.

Want one of these in your business?

Whether it's an off-the-shelf product or a custom build on our engines, we'll help you deploy it.

Get in Touch