Venture 02 · Engyn
Turn AI from a tool into an operating capability.
Engyn coordinates agents, skills, workflows, memory, governance and human oversight so intelligent systems can perform meaningful work across an organisation.
- Agents
- Goals
- Context
- Workflows
- Governance
- Stage
- In build
- Portfolio bay
- Bay 02
- Focus area
- AI & Software Engineering
- Part of
- The Collective Foundry
The problem
Why Engyn exists.
AI is easy to demo and hard to operate. Turning models into dependable work means coordinating agents, context, workflows and human oversight — usually stitched together by hand, with governance bolted on afterward.
Who it's for
Organisations that want AI to perform real operational work, not just answer questions.
The opportunity
The market Engyn is built for.
Organisations are moving from AI experiments to AI in production, and discovering the hard part is not the model — it is coordinating agents, context, governance and human oversight into dependable work. That orchestration layer is where the next wave of enterprise software is being built.
Why now
The timing behind it.
- 01
Foundation models are capable enough to act, not just answer — but they need orchestration and guardrails to be trusted with real work.
- 02
Enterprises are past the demo phase and want measurable operational output, with governance built in.
- 03
Agent frameworks are fragmenting; teams need a coordinating layer, not another isolated tool.
What it does
The capabilities inside Engyn.
Orchestrate agents
Coordinate agents, goals and missions so intelligent systems pursue outcomes, not one-off prompts.
Context & memory
Give AI the organisational memory, context and knowledge it needs to act with judgement.
Govern & oversee
Human approval, controls and governance built into every workflow from the start.
Operate & observe
Workflow execution and operational intelligence that make AI a measurable part of how work gets done.
What makes it innovative
Why Engyn is different.
Outcomes, not prompts
Goals and missions coordinate agents toward results instead of one-shot requests.
Governance-first
Human approval and controls are part of the workflow, not a compliance afterthought.
Compounding memory
Organisational context and memory make every workflow smarter than the last.
The usual way
Point AI tools stitched together by hand, with oversight bolted on afterward.
The foundry way
One orchestration layer where agents, context, governance and humans operate together.
How it compares
Engyn against the agent platforms an enterprise would already be evaluating.
Every competitor cell below is drawn from that vendor's own official site on the date shown. Where a capability was not documented there, it is marked Not publicly verified rather than Not available — absence of evidence is not evidence of absence.
| Capability | Engyn In build — no released tier | Microsoft Copilot Studio Copilot Studio with Microsoft 365 and Power Platform | Salesforce Agentforce Agentforce 360 (Observability requires Data 360 licensing) | Vertex AI Agent Builder Agent Engine runtime, consumption-based (memory charged per store and retrieve) | LangGraph / LangSmith Open-source LangGraph with LangSmith Deployment (renamed from LangGraph Platform, October 2025) |
|---|---|---|---|---|---|
| Making agents work together Multi-agent coordination toward a goal The difference between a chat box and a system that pursues an outcome. Also the row this entire category has already solved. | Planned Depth 0/5 · target 4/5 Parity, deliberately. Every competitor in this shortlist scores 5/5 here; goals and missions over per-prompt tasks is table stakes now, not a differentiator. | Advanced Shipped Depth 5/5 Generative orchestration with a planner, plus connected agents that delegate to other agents carrying their own tools and knowledge | Advanced Shipped Depth 5/5 Atlas decomposes a prompt into tasks and proposes a plan, evaluating at each step; topics, instructions and actions with subagents visible in the builder | Advanced Shipped Depth 5/5 Agent Engine runtime with native Agent-to-Agent protocol support, deploying agents written in ADK, LangGraph or CrewAI | Advanced Shipped Depth 5/5 Graph-structured orchestration with durable execution — the implementation the platforms above deploy rather than replace |
| Making agents work together Connecting agents to real systems and tools An agent that cannot act on anything is a demo. The breadth and the safety of that connection are different questions. | Planned Depth 0/5 · target 4/5 Parity. Microsoft and Salesforce each bring a connector estate built over a decade, and matching that breadth is not a realistic target for a new entrant. | Advanced Shipped Depth 5/5 The Power Platform connector estate, with knowledge sources kept deliberately read-only and actions as a separate layer | Advanced Shipped Depth 5/5 Actions built from Flow, prompt templates or Apex — acting on the CRM record directly rather than through an integration layer | Included Shipped Depth 4/5 Enhanced tool governance shipped as a named capability; the connector estate is narrower than Microsoft’s or Salesforce’s | Included Shipped Depth 4/5 Broad ecosystem of tool bindings, with middleware able to check each tool call against a policy. Every connection is code you write and maintain |
| Giving them something to reason over Organisational context and durable memory Without memory that survives the session, every run starts from nothing and no agent accumulates judgement. | Planned Depth 0/5 · target 4/5 Parity with Memory Bank rather than leadership. Durable cross-session memory is now a GA feature of the hyperscalers, not an open problem. | Included Shipped Depth 4/5 Knowledge sources against approved repositories, with non-overlapping sources recommended per subagent. Microsoft notes you cannot force an agent to use a specific article on demand | Included Shipped Depth 4/5 Dynamic grounding and secure data retrieval through the Trust Layer, with conversation logs in Data 360. Strongest where the context is already Salesforce data | Advanced Shipped Depth 5/5 Memory Bank at GA for persistent cross-session recall, separate from thread-scoped context, readable and writable by each subagent | Included Shipped Depth 4/5 Checkpointers for thread-scoped state and stores for long-term cross-thread memory, with TTLs on the managed server. The store is infrastructure you choose and run |
| Keeping humans in charge Human approval gates on consequential actions Decides whether an agent can be trusted with anything that costs money or touches a customer. | Planned Depth 0/5 · target 5/5 One of three leadership targets. “Built in, not bolted on” is the product’s own claim, and this is the row where the incumbents are least uniform — Copilot Studio and LangGraph are strong, the other two are not. | Advanced Shipped Depth 5/5 Three documented tiers: runs without confirmation, in-conversation confirmation, and offline approval through a workflow. Human-in-the-loop is mandated for high-impact actions | Limited Shipped Depth 3/5 Topic scope and guardrails constrain what an agent may do, and Agent Script defines deterministic paths — but a documented tiered approval model like Copilot Studio’s was not found | Not publicly verified Depth 2/5 Approval gating is achievable in a deployed graph, but a documented platform-level approval model was not found in the Agent Builder material | Advanced Shipped Depth 5/5 The reference implementation: interrupt() pauses a run mid-execution, surfaces a value, and resumes exactly where it stopped once a human responds |
| Keeping humans in charge Identity, policy, audit and agent-fleet control An organisation cannot run agents it cannot inventory, permission or audit after the fact. | Planned Depth 0/5 · target 5/5 Leadership target, and the hardest to justify: Microsoft already issues every agent an Entra identity. The claim can only be that governance holds across clouds, not that it is deeper inside one. | Advanced Shipped Depth 5/5 The strongest in this shortlist: a unique Entra Agent ID per agent, Agent 365 as cross-environment control plane, version control, release pipelines and approval gates | Advanced Shipped Depth 5/5 Einstein Trust Layer: zero data retention, toxicity detection, secure data retrieval, dynamic grounding, plus user-defined and Salesforce-managed guardrails | Included Shipped Depth 4/5 Enhanced tool governance plus the surrounding Google Cloud IAM and audit estate; no per-agent identity construct comparable to Entra Agent ID was found | Limited Shipped Depth 2/5 A framework, not a control plane. Policy middleware exists; agent inventory, identity and org-wide audit are left to the operator |
| Running it for real Deterministic workflow beside the reasoning Most real work is a fixed process with judgement at a few points. A platform that can only reason is the wrong shape for it. | Planned Depth 0/5 · target 4/5 Parity. Power Automate and Flow are mature deterministic engines sitting directly beside the reasoning layer in their own platforms. | Advanced Shipped Depth 5/5 Power Automate sits directly beside the agent layer, so deterministic process and reasoning are the same purchase | Advanced Shipped Depth 5/5 Agent Script combines deterministic workflow with LLM reasoning explicitly, on top of Flow as a mature process engine | Limited Shipped Depth 3/5 A runtime rather than a process engine. Deterministic workflow is something you build on Cloud Run or bring with your framework | Included Shipped Depth 4/5 Durable execution with checkpointed resumption is a strong deterministic substrate, but it is expressed in code rather than in a process designer |
| Running it for real Tracing, cost and operational intelligence Agents fail probabilistically. Without a trace of what one actually did, nobody can tell a bug from a bad day. | Planned Depth 0/5 · target 4/5 Parity. LangSmith and the Unified Trace Viewer both already render a full execution tree with cost and latency per span. | Included Shipped Depth 4/5 Agent 365 gives visibility into inventory, permissions, behaviour and activity; per-run trace depth is less documented than the governance surface | Advanced Shipped Depth 5/5 Agentforce Observability GA since November 2025, with full-trace views, reusable test cases generated from real interactions and batch testing. Requires Data 360 licensing | Advanced Shipped Depth 5/5 Tracing across reasoning, tool selection and execution paths, the Unified Trace Viewer, and a production dashboard for tokens, latency, errors and tool calls | Advanced Shipped Depth 5/5 LangSmith captures traces, token usage, cost and latency per span, thread and subagent, rendering the full execution tree across every trajectory |
| Running it for real Independence from one cloud, model or data platform The row that separates this shortlist. Every full product here is also a commitment to its vendor’s stack. | Planned Depth 0/5 · target 5/5 The actual wedge. Nothing in this shortlist is both neutral and operable: the three full products each commit you to a cloud, and the neutral option is a framework an engineering team runs itself. | Limited Shipped Depth 2/5 Excellent inside Microsoft 365 and Entra, and inseparable from them. That is the trade, and for most of its buyers it is the right one | Limited Shipped Depth 2/5 The deepest CRM integration here and the tightest coupling. Its strength on every row above is a function of the data living in Salesforce | Limited Shipped Depth 3/5 Genuinely framework-neutral — it will host LangGraph or CrewAI — but the runtime, memory and billing are Google Cloud | Advanced Shipped Depth 5/5 Open source, self-hostable, model-agnostic, and deployable on Vertex. The genuinely neutral option in this shortlist — and the one that is not a product |
Who each tool is actually best for
- Engyn In build — no released tier
Not yet a buying option. Engyn is in build inside the foundry; the scope below is intent, not shipped capability. Six of its eight targets are parity rather than leadership, because this category is already well served.
- Microsoft Copilot Studio Copilot Studio with Microsoft 365 and Power Platform
Organisations already on Microsoft 365 and Entra. The governance story is the strongest here — every agent gets an Entra Agent ID, and Agent 365 inventories the fleet.
Source: Generative orchestration — a planner that keeps each step within policy and seeks approval for sensitive actions , Multi-agent patterns — connected agents with their own tools and knowledge, and the governance checks they require , Agent governance update — Agent 365 as control plane, Entra Agent ID, release pipelines and approval gates , Cloud Adoption Framework — mandated human-in-the-loop confirmation for high-impact actions, strict per-agent data boundaries
- Salesforce Agentforce Agentforce 360 (Observability requires Data 360 licensing)
Organisations whose operational truth lives in Salesforce. Agents act directly on the CRM record through Flow and Apex, which no neutral platform can match without integration work.
Source: Atlas Reasoning Engine — decomposes a prompt into tasks and proposes a plan, evaluating at each step , Agent Builder — canvas, Agent Script hybrid reasoning, guardrails, batch testing from real interactions , Agentforce 360 — Agentforce Observability GA November 2025, conversation logs stored in Data 360 , How Agentforce works — Einstein Trust Layer: zero data retention, toxicity detection, dynamic grounding
- Vertex AI Agent Builder Agent Engine runtime, consumption-based (memory charged per store and retrieve)
Teams that want a managed runtime without framework lock-in. Agent Engine deploys agents built in ADK, LangGraph or CrewAI, and speaks the Agent-to-Agent protocol natively.
Source: Vertex AI Agent Builder overview — monitoring, logging and tracing including the Unified Trace Viewer , Build and manage multi-system agents — A2A protocol support, deploy agents from any framework , Agent Engine sessions and Memory Bank reach general availability , Enhanced tool governance in Vertex AI Agent Builder
- LangGraph / LangSmith Open-source LangGraph with LangSmith Deployment (renamed from LangGraph Platform, October 2025)
Engineering teams that want full control and no vendor commitment. The reference implementation of durable, interruptible agent execution — and the framework several platforms above will happily host for you.
Source: Human-in-the-loop and interrupts — interrupt() pauses a run and resumes on a human response , Persistence — checkpointers for thread-scoped state, stores for long-term cross-thread memory , Why LangGraph Platform — durable execution, task queues, 1-click deploys; renamed LangSmith Deployment , LangGraph overview — graph-structured orchestration with durable execution
Considered and excluded
- CrewAI and the other agent frameworks — LangGraph is compared as the representative framework, as the most thoroughly documented on the rows that separate a framework from a platform — durable execution, interrupts and persistence. Google lists CrewAI alongside LangGraph as a framework its runtime will host, which is the clearest statement of where both sit relative to this comparison. Check current status
How this comparison is made
- Comparative claims follow ACCC guidance: they must be accurate, fair and supportable on reasonable grounds, judged on the overall impression rather than line by line.
- A competitor capability is only marked as present where its official site or documentation describes it. Where it was not documented the status is “Not publicly verified” — not “Not available”. This is a desk review of published material, not a hands-on benchmark.
- All four competitors score 5/5 on agent orchestration, and three of them on observability. Engyn is late here and the targets reflect it: six of eight are parity rather than leadership. A matrix showing a new entrant beating three hyperscalers on every row would not be believable, and would not be true.
- Copilot Studio, Agentforce and Vertex are each strong and each tie you to a cloud, an identity system and a data platform — the portability row is 2, 2 and 3. The one genuinely neutral option, LangGraph, scores 2 on governance because it is a framework an engineering team operates itself. Nothing here is both neutral and operable, and that is the position Engyn targets.
- Depth scores (0–5) are our own analytical assessment of how deep a capability runs. They are not vendor-published figures, and they are separate from the evidence status beside them.
- Engyn is in build. Its public address is an intro page, not a released product, so every one of its cells shows delivery “Planned” and an achieved depth of zero. A row moves to “In build” only when there is internal evidence behind it, and its achieved depth is then capped at 2.
- Targets are published only for our own column and only with a stated reason. Three are set at 5: portability, which is the wedge, and the two control rows where the product’s own claim is that oversight is built in rather than added afterwards. The governance target is the hardest of the three to justify, and its note says so.
The approach
How we build it.
Engyn is an AI-native execution and intelligence platform. It gives AI agents the context, coordination, controls and tools required to perform meaningful work.
It spans agent orchestration, goals and missions, skills and harnesses, organisational memory, workflow automation, software delivery and operational intelligence — with human approval controls and governance built in, not bolted on.
The outcome: AI that operates inside your organisation with the oversight your organisation requires.
The outcome
What changes when it's in place.
- AI moves from a tool to an operating capability.
- Automated work stays under human control.
- Intelligence compounds as organisational memory grows.
How it fits the foundry
A venture, not a department.
Engyn moves through the whole foundry — discovery, design, engineering, quality and launch — like every venture we build. It's most visible in AI & Software Engineering, where its business purpose is sharpest.
- Discover
- Define
- Design
- Build
- Assure
- Launch
- Learn
- Scale