Summary

OpenAI is positioning the Responses API as the default surface for new agentic applications while keeping Chat Completions alive for years. swyx framed the intended direction as a “strict superset” combining Chat Completions’ capabilities with Assistants’ tools and state. The Assistants API is targeted to sunset in the first half of 2026; OpenAI plans to add assistant-like objects, thread-like objects, Code Interpreter, async mode, and webhooks, and give users a full year to migrate.
The platform is packaging web search, managed retrieval, computer control, orchestration, and observability around a unified agent stack. Developers can combine several capabilities in “one API request,” while Romain said teams that need complete retrieval control can still build custom infrastructure.
Web search shows how much tools can move model performance: SimpleQA accuracy rises from 38% for GPT-4o to 90% for GPT-4o Search Preview. Responses exposes search as a composable tool; Chat Completions instead requires the search-specific model, so o1 cannot simply call it there. Combined with function calling and structured outputs, Nikunj called it “an API for the internet, almost.”
Responses stores state free for 30 days, improving debugging without forcing statefulness. Developers can set store: false, while stored responses expose prompts, tool calls, and configuration failures in the dashboard. swyx nevertheless invoked Corey Quinn’s warning: “Anything that can be used as a database will be used as a database.”
File Search is positioned as managed RAG for teams that do not need to own every retrieval decision. It handles parsing, chunking, embeddings, vector storage, and search; the discussion also covered query optimization and custom reranking, while metadata filtering was identified as an especially important forthcoming feature beyond roughly 5,000–10,000 records. Full control still requires a hand-rolled stack.
Computer use remains an early, specialized capability rather than a mature general primitive. Its model consumes screenshots and returns clicks, scrolling, typing, and other tool calls across tasks that might require 20 steps and several minutes. Romain characterized it as “the GPT-2 of computer use or maybe GPT-1.”
The Agents SDK turns the unexpectedly popular Swarm experiment into OpenAI’s orchestration layer. Types, parallel guardrails, handoffs, tracing, and provider portability support specialized agents instead of “one agent to rule them all.” Longer term, OpenAI wants traces to generate evals and those evals to drive reinforcement fine-tuning, though “lots of details” remain unresolved.

Deep dive

1. Responses becomes OpenAI’s unified agent primitive

The launch bundles three built-in tools—Web Search, improved File Search, and Computer Use—with the new Responses API and an upgraded Agents SDK derived from Swarm.
swyx’s “strict superset” framing described the intended direction: Responses supports everything in Chat Completions at launch, is intended to absorb Assistants capabilities over time, and can remain stateless through store: false.
Chat Completions, originally launched around March 2023 for single-turn text workflows, is “here to stay” with new models and features. Assistants instead has a target sunset in the first half of 2026; OpenAI plans to add assistant-like objects, thread-like objects, Code Interpreter, async mode, and webhooks to Responses and give users a full year to migrate.
State is stored free for 30 days. Alessio highlighted dashboard-level debugging—seeing the prompt, tool choice, and configuration—while swyx quoted Corey Quinn’s warning about anything usable as a database; Romain noted that users already stuff data into object metadata.

2. Web Search turns live information into structured application data

Search ships two ways: as a built-in Responses tool, or as GPT-4o Search Preview inside Chat Completions. The latter cannot be attached as a tool to another model such as o1.
The search fine-tune focuses on retrieving the right material, remaining factual, and citing accurately; Nikunj cited synthetic-data techniques and model distillation. On SimpleQA, the hosts highlighted 90% accuracy versus GPT-4o’s 38%.
Nikunj’s strongest composability example combined Web Search, function calling, and structured outputs: live web information can arrive directly in an application’s required JSON schema—“like an API for the internet, almost.”
The hosts pushed on retrieval depth and cost: top-K can return irrelevant filler, while a similarity cutoff might return five or 500 documents and make spending unpredictable at roughly $30 per 1,000 queries. Nikunj said depth is not a parameter today and suggested orchestration with a planning step and progressively deeper calls; swyx proposed a context budget as another way to manage cost.

3. File Search offers managed RAG, with control as the trade-off

File Search accepts private data and manages parsing, chunking, embeddings, search, and the vector store. Query optimization and custom reranking were also discussed, while metadata filtering was identified as the main requested addition and especially important once a store exceeds about 5,000–10,000 records.
Nikunj’s best combined-tool example stores a user’s reading or fashion preferences in File Search, retrieves them as memory, then searches the web for matching products—all within one Responses call: configure the tools, “go boom and like everything just happens.”
swyx cited Navan loading FAQs and travel policies into File Search so an assistant can use those policies without rebuilding retrieval infrastructure. OpenAI’s recommendation was pragmatic: start managed, then hand-roll chunking and retrieval only when complete control materially matters.

4. Computer Use extends agents into long-horizon interface work

The computer-use model behind Operator receives screenshots and usually returns tool calls—click, scroll, type, and report back. A browser task may take “a couple minutes” and perhaps 20 steps, explaining why Responses was designed for multi-turn, longer-horizon execution.
Romain called the capability “the GPT-2 of computer use or maybe GPT-1.” The hosts proposed Pokémon as an agent benchmark; OpenAI agreed it was worth trying, while Romain said the vision is for preview fine-tunes to merge into core models once stable, as vision previously did.

5. The Agents SDK makes handoffs observable and trainable

Swarm began as a “low-key experiment” for multi-agent orchestration, but unexpected adoption persuaded OpenAI to promote handoffs into a core platform primitive under the Agents SDK name.
The SDK adds types, guardrails that can run in parallel and block execution, and built-in tracing. Responses is the default, but any provider implementing the Chat Completions format can plug in, as can multiple tracing providers.
Romain contrasted a monitored triage agent handing work to specialists with “one agent to rule them all” stuffed with tool calls. Dashboard traces show each handoff, downstream agent, and tool call for troubleshooting.
Romain’s roadmap connects traces to eval generation, then uses good evaluators and tasks for reinforcement fine-tuning. “Lots of details” remain, but OpenAI intends to pursue that trace-to-eval-to-RFT loop “pretty hard.”

Episode 133 - March 7, 2025 Alex Morris, TSOH & Author of "Buffett and Munger Unscripted", talks 2024 Berkshire Annual Letter