Building more effective AI agents
- Source
- youtube.com
- Author
- Anthropic
- Date
Context
Chaining agents and running agents in parallel are different designs, according to Erik, who works on research at Anthropic, in an Anthropic video. In a "workflow of agents," one agent finishes and its output goes to the next step. The example is an application whose SQL step becomes a loop that runs the query, checks the result and retries until the value looks right, before a charting step begins. In a multi-agent system, a parent agent delegates to sub-agents that work at the same time.
Erik says Anthropic's deep research product works the second way: an orchestrator creates sub-agents that run many searches in parallel, so the user gets an answer sooner. Claude Code can also hand a bulky sub-task, such as finding a class implementation, to a so only a short answer returns to the main . These are the speaker's descriptions of Anthropic's own products, not independent tests. Erik also advises starting with the simplest setup, because observability gets harder with each layer. Related Intel: Anthropic's agent-building tips.
- Erik says agent loops now beat fixed workflows when quality matters most, while workflows still suit low-latency single-shot answers. He describes workflows of agents, where each step, such as writing a SQL query, loops until the output is right.
- Erik explains that to Claude a subagent looks like a tool it passes a prompt to. Subagents shield the main context from token-heavy subtasks, but Claude makes first-time-manager mistakes like unclear instructions and grows more detailed with training.
- Erik reports that customers with 100 or 200 tools found it works well to split them among subagents, so the main agent only picks a bucket of tools and each subagent needs to understand maybe 20.
- Erik argues tools and MCP servers should map to your UI, not your API. If Slack is exposed as separate endpoints for messages, user IDs and channel IDs, the model needs three calls to understand anything; one tool should present it all at once.
- Erik warns that overbuilt multi-agent systems can spend too much time talking to each other instead of making progress. He advises starting simple, adding complexity only as needed, and remembering the model sees only what you showed it.
- multi-agent — Using several AI agents on one problem — splitting work in parallel, checking each other, or filling different roles like planner and reviewer.
- subagent — A separate agent spawned by another to do one scoped piece of work in its own context, returning only the result.
- context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
- test-time compute — Spending more computation when the model answers — thinking longer, trying multiple attempts — to buy accuracy without training a bigger model.
Checking sign-in…
Loading comments…






