know.2nth.ai › Agents › Agents, Data & Trust

agents · Perspective · Beyond the Hype

Agents don't expose your engineers. They expose your data.

In a fireside chat at LangChain's Interrupt conference, Andrew Ng and Harrison Chase moved the agent conversation off chatbot hype and onto the structural reality: when coding agents make building cheap and fast, the bottleneck doesn't disappear — it climbs the pipeline to product scoping and data architecture. This node summarises that argument and validates its named open-source toolchain against primary sources, with licences stated precisely. It's the thesis that ties the rest of the agents tree together — and the case for getting your data layer right first.

Ng × Chase MCP through-line Fact-validated Data & trust Anti-hype

01 · What it is

The shift from the hype era to the leverage era.

A fireside chat between Andrew Ng (founder of DeepLearning.AI and AI Fund) and Harrison Chase (co-founder and CEO of LangChain) at Interrupt, LangChain's agent-focused conference. The conversation is a deliberate move away from consumer chatbot hype and existential-risk narratives toward the structural realities of putting agents to work inside a business.

The core argument: the interesting shift isn't standalone AI tools or job-loss panic — it's coding and workflow agents that compress execution timelines, and what that compression exposes downstream. When building software gets cheap and fast, the constraint stops being the engineer and moves up the pipeline to product scoping and data architecture.

Hype era                     →    Leverage era
Standalone chatbot / "tool"      System-wide scaling engine
Engineer = bottleneck            PM + data architecture = bottleneck
Automate one task                Combine building blocks (compounding)

Primary source & validation discipline

The full session is at youtu.be/OaRhpwz_TGM ↗. The cited video is the 2026 Interrupt edition ("The Future of AI Agents"), distinct from the widely-circulated 2025 "State of AI Agents" session, so claims here are attributed to the Ng/Chase fireside generally rather than pinned to one year's transcript. Every tool named below is validated against its primary GitHub repository, with the licence stated — because "open source" gets flattened too often.

02 · Why it matters

Where the rate-limiting step moves.

The headline isn't "agents are coming." It's where the bottleneck goes once they arrive. Five claims from the talk, restated plainly.

Coding agents matured quietly. While public attention sat on existential risk, engineering workflows moved from prompt-response to multi-tool coding agents. Engineers increasingly act as supervisors — launching and reviewing agent tasks rather than typing every line, sometimes from a phone.
The bottleneck moved to product management. If an agent can build in an afternoon what took three weeks, the developer is no longer the limit. The organisation now has to know what to build and scope it precisely. Speed without clear business logic just produces faster chaos.
Cost-cutting is the small prize. Deployments that only trim budgets or shave task times (faster loan underwriting, say) miss the larger move: applying AI to top-line growth and combinatorial capability-building.
Data architecture is the prerequisite. Most enterprise data sits in fragmented, unstructured silos. Useful corporate agents need a rearchitecture of unstructured data — centralised "context" that keeps agents grounded in the actual state of the business.
Avoid lock-in. Because the landscape shifts fast, keep the stack modular — open-source models, open frameworks, and observability — so you can hot-swap models as better ones appear.

Task automation vs skill-building

This maps directly onto the move from task automation to skill-building. Automating one isolated task gives an isolated gain; an integrated, data-grounded agentic workflow compounds. The value isn't any single automation — it's the composition.

03 · The argument in four moves

Cheap build → exposed scoping → grounded data → compounding.

Move 1

Coding agents become team members

Code generation stops being autocomplete and becomes an agentic loop: read files, plan, edit, run commands, run tests, iterate with human approval. The human shifts from author to reviewer.

Move 2

The constraint climbs the pipeline

Compress the build step and whatever sits above it becomes binding: product scoping and validation. Precise requirements, clear role boundaries for "digital employees", fast product-market fit.

Move 3

Combinatorial leverage

AI capabilities as LEGO bricks — reusable building blocks that combine into net-new capabilities rather than one-off automations. The value is in composition, not any single block.

Move 4

Ground the agents in real data

Before scaling an agentic workforce, fix the data pipeline. Agents need continuous, structured access to the business's real state to stay safe and accurate. Without it, they're confidently wrong.

Cheap build step  →  exposes  →  PM / scoping bottleneck
                                        │
                  requires  →  grounded, structured data
                                        │
                  enables   →  combinatorial capability (compounding)

04 · The validated toolchain

Named tools, real licences, MCP as the filter.

The talk references a shift from single-platform coding tools toward an integrated ecosystem of agents. Here are the named open-source components, validated against their primary GitHub sources, with licences stated precisely.

Tool	Licence (validated)	Agent / MCP surface
LangChain	MIT (core)	Framework; 1000+ integrations
LangGraph	MIT (core)	Low-level orchestration runtime; durable execution, human-in-the-loop
OpenCode	MIT (you pay for models)	Terminal coding agent; client/server (drive from mobile); MCP support
Gemini CLI	Apache-2.0	Terminal coding agent; native MCP server support; built-in tools
LangSmith	Proprietary / freemium	Observability, tracing, evals — framework-flexible but vendor-owned

Two honest flags from the validation log

LangSmith is not "vendor-neutral". It is model- and framework-flexible, but it's LangChain's own commercial product. If true neutrality is the requirement, OpenTelemetry-based observability is the open alternative.
OpenCode naming collision. The original repo has been mirrored and forked widely under the same name; pin to the project's own domain (opencode.ai) to identify the canonical build before relying on any fork.

MCP — the through-line

Both OpenCode and Gemini CLI expose Model Context Protocol support, which is precisely the "agent-driveable surface" filter that matters: it's how an agent gets standardised, swappable access to tools and data rather than bespoke glue per integration. MCP is the seam where the "grounded data" prerequisite actually meets the agent — which is why it sits at the centre of this whole tree.

05 · Use cases

What the thesis looks like in practice.

Supervised coding workflows — engineers launch agent tasks (build, refactor, test) and review diffs, including from mobile, instead of hand-writing every change.
Combinatorial product builds — composing reusable agent "bricks" into net-new internal capabilities rather than automating one task at a time.
Context hubs — centralising fragmented unstructured data so agents stay grounded in the live state of the business.
Provider-agnostic stacks — building on open frameworks so the underlying model can be swapped as better or cheaper options appear.
The "10-person powerhouse" — small, AI-fluent specialist teams shipping platforms that previously needed far larger headcount (a claim from the talk, not a measured figure).
Top-line applications — moving AI from cost-trimming to growth: new services, previously unviable markets.

06 · Decision guide

Use this thesis when. Fix first when.

Use this thesis when

You can already scope a product precisely, and build speed is the real constraint
Your data is reachable and reasonably structured — or you'll invest to fix it first
You want a modular stack and are wary of single-vendor lock-in
Managers can define role boundaries for digital workers and validate fit fast

Skip / fix first when

Data sits in fragmented silos with no access layer — agents will be confidently wrong
The real bottleneck is unclear requirements, not engineering throughput
You're treating AI purely as a cost-cutting line item — the compounding value is on growth
You need true vendor-neutral tooling and can't accept a commercial layer in the critical path

Anti-hype note

"Compress three weeks into an afternoon", "2ⁿ scale" and the "10-person powerhouse" are rhetorical framing from the speakers, not benchmarks. Treat them as direction, not a service-level promise. Time-bound popularity metrics (star counts, monthly-user figures) are excluded by policy.

07 · South African context

"AI readiness" is an audit, not a purchase.

In SA delivery, the thesis lands with an extra constraint: the data rearchitecture — that centralised "context hub" — has to account for where the data physically lives.

The prerequisite is the data layer — which is the back office

"Grounded, structured data" is exactly the back office below the waterline — ERP, records, suppliers, finance. Agents are the visible payoff; the owned, reachable data layer is what makes them safe. That's why this thesis and the ownership story are the same argument seen from two ends: you can't point an agent at a black-box vendor silo and expect it to be grounded.

POPIA · where the context hub physically lives

Centralising business context for agents raises an immediate SA question: where does that hub live, and what personal data does it hold? POPIA data-residency makes the location of the context hub an architectural decision, not an afterthought — keep residency-sensitive context in-country and govern what reaches an offshore model. The edge is where that border gets enforced.

The honest sequencing

Before scaling an agentic workforce: assess data cleanliness and accessibility, define role boundaries for digital employees, and equip managers to clear the product-scoping bottleneck. The data rearchitecture is the gating investment. Scale agents onto a clean data layer — not before it.

08 · Where this gets done

From the thesis to a ready data layer.

AI readiness · the engagement

The audit before the agents

This knowledge tree explains the thesis. Doing it — the AI-readiness audit, the data rearchitecture that builds a grounded context hub, defining role boundaries for digital workers, and the POPIA-aware decision on where that context lives — is the engagement Imbila runs. It's the gating investment that turns "agents are coming" into agents that are actually grounded, safe and compounding.

Imbila — AI strategy & implementation ↗

09 · Connections