Harrison Chase on Production Agents #4: Deep Agents

Part 4 of 5 — tracing Chase's production agent thinking with FRE|Nxt Labs commentary.

---

The launch

"Deep agents come with batteries included — planning, memory, subagents, file systems, and starter prompts — so you can just start building."

Deep Agents implement a combination of four things: a planning tool, sub-agents, access to a file system, and a detailed prompt. Applications like Deep Research, Manus, and Claude Code have gotten around the limitation of "shallow" agents by implementing these characteristics.

— LangChain — Deep Agents, launched July 2025

Chase's own summary from a related thread: "agents are a filesystem. memory built in. triggers turn it into an ambient agent. a feed for reviewing and approving agent work. supports MCPs, skills, and subagents."

---

What we heard

Deep Agents is Chase packaging the production checklist from entry #3 into reusable infrastructure. Every production agent we've shipped manually implements something like this: a planner, a memory layer, delegable subagents, and a virtual filesystem. Deep Agents makes those first-class in the framework.

The specific design decisions are worth reading:

• Planning as a tool, not a pipeline: the LLM calls a todo-list tool dynamically instead of running a scripted plan. This matters because agent work is discovered, not pre-specified.
• Filesystem as state representation: the agent offloads memory to files instead of holding everything in the context window. Treats context like RAM; treats filesystem like disk. (Echoes Karpathy's LLM OS framing.)
• Subagents as specialists: each subagent has its own tools, config, and prompt. The coordinator delegates specific, bounded tasks. This is how Claude Code, Deep Research, and similar products actually work in production.

---

What we actually do with this

On every new engagement we ask a build-vs-adopt question: does Deep Agents (or a similar packaged harness) already solve 80% of what this system needs?

Our decision tree:

| Signal | Action | |---|---| | System needs a coordinator + specialist subagents with shared state | Strongly consider adopting Deep Agents or similar | | System has a single agent doing many tasks in one prompt | Refactor into Deep Agents subagent pattern regardless of framework choice | | System's failure modes are task-planning-related | Deep Agents' explicit planning tool will likely help | | System's state is simple (single conversation, no long-horizon work) | Deep Agents is overkill — plain LangGraph is fine | | Client has strong preferences or existing framework | Respect the preference; reach for the Deep Agents patterns even if not the library |

Importantly: we don't adopt Deep Agents reflexively. The packaged harness is powerful when you need what it provides. When you don't need planning-as-tool or a filesystem abstraction, the extra abstraction makes the system harder to reason about.

---

Applied: where we used the Deep Agents pattern (not the library)

On InterviewLM we did not adopt Deep Agents as a library (we shipped before its public launch). But we reached for three of its core patterns independently:

1. Coordinator + specialist subagents: an interview coordinator routes to persona-specialized subagents (technical, behavioural, case study). Each subagent has its own prompt, rubric access, and evaluation scope. Exactly the Deep Agents pattern, implemented in LangGraph.
2. Filesystem as state: the transcript, rubric, and running evaluation notes live in structured files (versioned in the session store). The subagents read/write these files rather than holding full history in context. This is the Deep Agents "agents as filesystem" idea, applied.
3. Planning as a callable tool: the coordinator has a plan_next_phase(session_state) tool that returns a list of subagent invocations for the next 1–3 turns. We explicitly rejected a scripted pipeline because interview progression is adaptive.

That we arrived at the Deep Agents pattern independently is itself evidence that Chase packaged the right primitives.

---

When not to use Deep Agents

A common failure we see in 2026: teams adopt Deep Agents for single-agent use cases where a LangGraph node or even a plain LangChain composition would do. The extra abstraction is a net cost when the problem doesn't need it. Deep Agents is a harness for multi-step, multi-specialist work. Don't reach for it when you have one agent doing one job.

---

The one thing to steal from this

Audit your current agent system against the four Deep Agents primitives: planning tool, subagents, filesystem, detailed prompt. Which are you missing? Missing planning → agent can't self-adjust. Missing subagents → one prompt is doing too many jobs. Missing filesystem → context window is your bottleneck. Missing detailed prompt → you're leaving capability on the table. Add whichever one is your weakest.

---

Next in this series

#5 — Ambient Agents (2025–26). Chase's vision for the next UX: agents that act on event streams, not chat prompts — and why "ambient" doesn't mean "autonomous."