Every reliable agent is built from the same handful of patterns

Blocks is our library of production-tested agentic patterns - planning, reflection, tool use, ReAct, orchestration, human-in-the-loop. Compose them into an agent instead of hand-writing a reasoning loop for every new project.

6 core patterns
Composable & model-agnostic
Governed by Sentinel
blocks.library
6 patterns
Planning
decompose
Reflection
self-critique
Tool use
call systems
ReAct
reason ↔ act
Sequential
agent pipeline
Human-in-loop
approval gate
6 patterns
·
composed by Forge
·
governed by Sentinel
Explore
The bottleneck

Every agent reinvents the same reasoning loop

Planning, reflection, tool use, a ReAct loop, an approval gate - the same handful of patterns sit inside almost every agent worth shipping. Most teams hand-write them from scratch each time, and rebuild the same bugs along the way.

Every team reinvents the loop

Each new agent starts with someone re-implementing planning, retries, tool-calling and a reason-act loop by hand. The interesting, domain-specific part ships buried under scaffolding a hundred teams have already written a hundred times.

Brittle agents that demo, then fail

The loop that looked flawless on a happy-path demo has no reflection step, no escalation gate and no test suite behind it. It meets a real edge case in week two, and there's nothing underneath to catch the fall.

No shared, tested primitives

Because every agent's internals are bespoke, nothing is reusable, nothing is evaluated the same way twice, and governance is applied - or forgotten - case by case. There's no library to build on, only a growing pile of one-offs.

What Blocks does

Six patterns, tested and ready to compose

Blocks packages the agentic patterns worth reusing - each one configured to your models and tools, tested with eval cases, and governed the same way underneath. Compose them; don't rewrite them.

Planning

Decompose a goal into an ordered set of steps, then execute them. The agent writes the plan before it acts, so a complex task becomes a sequence you can inspect, resume and audit - not one opaque prompt.

DecomposeOrdered stepsResumable

Reflection

The agent drafts, critiques its own output against your criteria, then revises before returning. Self-correction catches the mistakes a single pass makes - the missing filter, the wrong unit, the unstated assumption.

DraftSelf-critiqueRevise

Tool use

Reason about which system to call, call it through MCP, a REST API or a function, and use the result. Blocks handles the tool contract, the arguments and the observation, so the agent acts on live data, not a stale guess.

MCPREST & functionsLive observation

ReAct loops

Interleave reasoning and action in a loop - think, act, observe, repeat - until the task is solved. It's how an agent works through a problem it can't answer in one shot, one grounded step at a time.

ThinkActObserveRepeat

Sequential & multi-agent

Chain specialized agents into a pipeline, each handing its output to the next - extract, then validate, then summarize. Specialists in sequence beat one generalist doing everything at once.

ExtractValidateSummarize

Human-in-the-loop

Bring a person in exactly where the stakes require it - an approval, a review, an escalation. Set the condition, and the agent pauses for a human decision that's captured to the audit trail.

Approval gateEscalationLogged decision
See it work

Explore the pattern library

Pick a pattern, read how it's structured, then play its execution trace step by step. This is an illustration of how each Block reasons and acts - not a live model call.

Planning
Reflection
Tool use
ReAct
Sequential
Human-in-the-loop
Planning
Agentic pattern
When to use

Break a big goal into ordered steps, then execute them.

Structure
Good fit for
Composable
Model-agnostic
Governed by Sentinel
Trace
About
Idle
Play trace

Press Play trace to stream this pattern's reasoning and actions, step by step.

0 steps traced
pattern: Planning

Six patterns, one governed runtime - compose them into an agent, or let Forge compose them for you.

Under the hood

Six steps from library to a running agent

A Block goes from the library to production the same way every time - configured, composed, evaluated and governed before it ever ships.

1

Pick a pattern

Choose the pattern that fits the job - planning, ReAct, reflection or another. Start from a tested primitive instead of a blank file.

2

Configure

Wire in your model, your tools and your prompt. The pattern's structure stays fixed; what it reasons over is entirely yours.

3

Compose

Nest and chain patterns into the agent you actually need - a ReAct loop that uses tools, wrapped in reflection, gated by a human step.

4

Evaluate

Run the composed agent against your test cases. Blocks are built to be measured, so reliability is a number you can see, not a hope.

5

Govern

Sentinel applies guardrails, eval gates and an audit trail to every pattern and every run - governance as a property of the Block.

6

Ship

Deploy inside Forge, in your own cloud - as an API, a job or a step in a larger workflow. Versioned, observable and yours to change.

Reference architecture

A pattern library the Forge runtime composes

Blocks isn't a model or a product - it's the layer of proven patterns underneath them. You compose from the library; Forge runs the composition over your models, tools and data; Sentinel governs the whole thing.

Compose
Blocks - the pattern libraryplanning · reflection · tool use · ReAct · sequential · human-in-the-loop
Pattern composernest · chain · configure
Forge runtimestate · memory · execution
Eval & registrytest cases · models · tools · MCP
Your modelsClaude · open-weight · fine-tuned
Your tools & APIsMCP · REST · internal systems
Your dataDatabricks · warehouse · vector
Deploy targetsAPI endpoint · scheduled job · a step inside a Forge workflow - in your cloud
Governed end to end - by Sentinel
Guardrails
Evaluation gates
Full-run audit trail
Access control
Observability
Run

Deployed inside your cloud and your security perimeter - your patterns, your data and your logic never have to leave to run.

Composition

How Blocks compose

Patterns aren’t used alone - they nest and they stack. A tool call sits inside a ReAct loop; the loop is wrapped in reflection; the whole thing is gated by a human step. The same primitives compose upward into Forge workflows and the products your users see.

They nest
Patterns wrap each other
Human-in-the-loop
Reflection
ReAct loop
Tool use
Model call
the task itself

A tool call inside a ReAct loop, wrapped in reflection, gated by a human approval - one governed agent, composed from four patterns.

They stack
Primitives compose up into products
Solutions
copilots · document & finance agents · research pipelines
Forge
composes & runs the patterns · state · retries
Blocks
the six agentic patterns
Planning
Reflection
Tool use
ReAct
Sequential
Human
Governed by Sentinel at every layer
In production

The patterns behind agents we already run

Blocks is the library underneath the multi-agent systems Focaloid ships today - the reason a reliable agent is composed in weeks, not invented from scratch.

Security · CVE research
In production

A multi-agent CVE research pipeline

Thousands of CVEs, triaged by a pipeline of specialized agents - one plans the investigation, others use tools to pull advisories and match affected versions, a sequential hand-off assembles the assessment. Triage that took analysts hours now runs in minutes. It’s Planning, Tool use and Sequential orchestration, composed from the library.

Where teams point it next

The same patterns, pointed at the next problem

The identical building blocks sit behind the media-monitoring and NLP agents that read and summarize at scale, and the document and finance agents that extract, validate and reconcile. Different domain, same tested primitives underneath - composed to each.

Where it fits

A pattern for each kind of job

Reach for the pattern that matches the shape of the work - and remember that most real agents combine a few of them.

Planning → research & reports

Multi-step research, board summaries, analyses - anywhere the agent needs a plan before it acts and a result you can trace back to its steps.

Board reportsMarket researchDue diligence

ReAct → lookups & troubleshooting

Investigations, diagnostics and data lookups where the agent has to think, query, read the result and decide the next move, one grounded step at a time.

TriageRoot causeData lookups

Reflection → drafting & QA

Drafting, code and content generation, quality review - anywhere a self-critique-and-revise pass turns a rough first answer into one you’d actually send.

Copy & codeQA reviewSummaries

Tool use → system actions

Order status, account updates, live data - whenever the agent has to reach into a real system through MCP or an API rather than reason in a vacuum.

Order statusAccount updatesTicketing

Sequential → multi-step pipelines

Document processing, data enrichment, back-office flows - work that’s naturally a chain of specialized stages, each handing off to the next.

ClaimsOnboardingReconciliation

Human-in-the-loop → regulated decisions

Refunds, approvals, underwriting, anything over a threshold - decisions that need a person’s sign-off, captured with the evidence for the audit that follows.

RefundsUnderwritingCredit limits
Why Blocks, why us

Building blocks you can stake a regulated process on

There are plenty of ways to sketch an agent loop. Fewer are built so you’d run a regulated process on the result - and hand it to your own team afterward.

Proven, not experimental

Every pattern is one we run in production, not a diagram from a paper. You start from primitives that have already met real edge cases and survived them.

Composable & model-agnostic

Nest and chain patterns freely, and swap the model underneath - Claude, open-weight, fine-tuned - without touching the structure that makes them reliable.

Governed by design

Guardrails, eval gates and a full audit trail come standard through Sentinel, applied to every pattern and every run - not bolted on when compliance asks.

Tested with eval gates

Blocks are built to be measured. Reliability is a number from your own test cases, checked before deploy, not a claim from a one-off demo.

Yours to keep

The patterns run in your cloud, composed inside Forge, defined in your repo. No dependence on us to change an agent once it’s live.

One partner across the journey

From the first pattern to a governed multi-agent system in production, it’s one team - the same people who built Forge, Sentinel and Prism.

Solutions & Accelerators

Accelerators that make production faster - and safer.

Two kinds of reusable IP: the tooling we build and govern with, and the solutions that drop straight into a use case. Blocks is the first of the former - the layer everything else on this board is assembled from.

Blocks is the bottom of the stack. Forge composes and runs the patterns, Velocity uses them across the delivery lifecycle, Sentinel governs every one - and Prism and Echo are what they add up to when pointed at a real use case.

Trust & governance

Patterns ready for the review they’ll face

Blocks is built for regulated work. ISO/IEC 27001-certified delivery, governance aligned to US frameworks like the NIST AI RMF and the EU AI Act, an audit trail on every run, and deployment inside your own cloud. Every pattern is governed by Sentinel - so an agent composed from Blocks clears the review instead of stalling in it.

How Sentinel governs it
What’s inside

The technology behind Blocks

Proven orchestration frameworks, models and evaluation tooling - assembled into a library of patterns, chosen for reliability and fit rather than a logo on a slide.

Patterns & orchestration

LGLangGraph
CRCrewAI
RAReAct
PLPlanner

Models & reasoning

CLClaude
OWOpen-weight LLMs
FTFine-tuned models

Tools & integration

MCPMCP servers
APIREST & webhooks
FNFunction calling

Eval & quality

EVEval harness
TCTest cases
TRTracing

Data & platform

DBDatabricks
UCUnity Catalog
VECVector store

Governance

SENSentinel guardrails
AUDAudit trail
OBSObservability

The exact stack is chosen per engagement - Blocks is framework- and model-agnostic by design, so it can adopt a better pattern or model the day one arrives.

Questions we get

Before you build on Blocks

Is Blocks a product we buy, or a library?

A library, delivered as part of how we build. Blocks is our collection of production-tested agentic patterns; we compose them into your agents - usually inside Forge - and hand the result over. You don’t buy a box; you get agents built from primitives you can keep and change.

How is Blocks different from just using LangGraph or CrewAI?

Those are excellent frameworks, and Blocks uses them. The difference is what sits on top: patterns we’ve already hardened in production, configured to your stack, tested with eval gates and governed by Sentinel. You get the framework’s power without re-deriving the reliable way to use it for every new project.

Which models does it work with?

Any of them. Blocks is model-agnostic - Claude, open-weight, or your own fine-tuned models - and a pattern’s structure doesn’t change when you swap the model underneath. Use the right model for each step, and change it later without a rewrite.

How do you keep agents reliable?

Every pattern is built to be evaluated. We baseline a composed agent against your real test cases, put eval gates in front of deploy, and trace every run - so reliability is measured before anything ships and watched after it does, rather than assumed.

How does governance actually work?

Every Block runs under Sentinel: guardrails on inputs and outputs, evaluation gates before deploy, access control and a full audit trail on every run. Governance is a property of the pattern, not a checklist someone remembers to apply.

How do we adopt it?

Through Forge. We compose the patterns you need into a workflow, run it in your cloud, and give your team the visual canvas and the definition to change it. Blocks is the library; Forge is how you build and run with it. A focused first agent is a matter of weeks.

Have an agent worth building on solid ground?

Bring us the task - a research pipeline, an approval flow, a document agent. On a 30-minute call we’ll map it onto the patterns it’s really made of, and show you what it takes to compose, govern and ship it.

ISO/IEC 27001-certified · Governed by design · Deploys in your cloud