When to Use DeerFlow: A Decision Model for Durable Agentic Work

DeerFlow 2.0 is worth using when a language model must carry out a meaningful workflow rather than merely produce a one-off answer. This final synthesis asks when its surrounding Gateway, lead agent, capabilities, sandbox, continuity, coordination, and evidence mechanisms create enough value to justify their cost and operational demands. The answer is conditional: DeerFlow is strongest where work has bounded side effects, needs continuity or coordination, and benefits from an inspectable path from objective to deliverable.

A harness is a contract around a model

The most concise description of DeerFlow 2.0 is an execution control plane for model-driven work. The model supplies flexible interpretation and next-action judgment. DeerFlow supplies the conditions under which that judgment becomes an operation: a user and thread identity, a set of visible and authorized capabilities, a workspace or sandbox, structured state, continuity mechanisms, execution limits, event streams, and a terminal record.

Those conditions form a contract rather than a guarantee. The Gateway owns the run lifecycle and connects the user-facing application to the embedded LangGraph-compatible runtime. The lead agent interprets the objective and selects actions. Middleware mediates the exchange, applying input and remote-content boundaries, authorization, sandbox access, context handling, tool-result limits, delegation policy, and terminal recovery. Skills supply workflow knowledge and, after activation or capture, can narrow the tools that remain appropriate. MCP and other providers extend the capability surface, but they also introduce remote authority, credentials, availability, and protocol assumptions.

The thread gives the work a scope. Uploads, intermediate files, output artifacts, structured state, and selected continuity belong to a user and conversation rather than floating in a global process. The sandbox determines where commands and file operations occur. A local provider may be convenient but is not a secure host boundary for shell access. A container or remote provider can offer stronger isolation, at the price of startup, synchronization, networking, capacity, and ownership complexity.

Persistence then answers a more specific question than “does the agent have memory?” Checkpoints preserve the state of an ongoing graph and its thread. Summaries compress old conversation. Durable context preserves selected delegation and skill references so compaction does not erase the runtime facts needed for the next request. Long-term memory stores selected user or agent knowledge across sessions. Goals express a completion condition that can produce bounded continuations. Scheduled and remote-task services preserve work that outlives the immediate model call.

Finally, evidence makes the contract inspectable. Stream events show what is happening now. Run events and receipts preserve bounded records of actions. Workspace snapshots show what changed. Delivery checks connect produced output files to the result presented to the user. None of these layers turns model judgment into verified truth. Their value is that they separate questions which a simple chat transcript cannot answer: what was available, what was authorized, what ran, what changed, what was delivered, and what still needs substantive review.

erDiagram USER ||--o{ THREAD : owns OBJECTIVE ||--o{ RUN : initiates THREAD ||--o{ RUN : contains RUN }o--|| LEAD_AGENT : orchestrated_by LEAD_AGENT }o--o{ CAPABILITY : invokes THREAD ||--|| WORKSPACE_OR_SANDBOX : scopes WORKSPACE_OR_SANDBOX ||--o{ ARTIFACT : yields THREAD ||--o{ CONTINUITY_STATE : preserves USER ||--o{ MEMORY : accumulates RUN ||--o{ SUBAGENT_TASK : may_delegate RUN ||--o{ EVIDENCE : records EVIDENCE }o--|| ARTIFACT : may_verify

The relationships show why DeerFlow is more than a model wrapper. An objective initiates runs, but a run lives inside a thread. The lead agent invokes capabilities under policy, and those capabilities operate through a thread-scoped workspace or sandbox. Continuity can preserve a thread’s current state or recall selected user knowledge, while sub-agent work remains a related but distinct execution. Evidence records the observable path and may verify that an artifact was delivered; it does not replace the human or domain-specific judgment of whether the artifact is good.

This model also explains the project’s move from Deep Research to a super-agent harness. Research is a particularly revealing workload because it combines browsing, synthesis, files, context, and a visible report.