October 2, 2026
Review without another agent's reasoning: Deep Agents gives subagents separate context
One AI agent searches for a solution while another checks it with a fresh context: Harrison Chase breaks down this technique from Cogentic.
Harrison Chase
@hwchase17
A good technique from this Google paper the verifier starts with a fresh context, so the agent searching for a solution cannot use its reasoning to persuade it to accept an invalid proof so the agent can try the boldest ideas, while the verifier rejects invalid ones this is also why subagents in Deep Agents have their own context https://x.com/omarsar0/status/2106056369816420624
Banger paper from Google Research on multi-agent proof discovery. (bookmark it) It's really interesting to see this emerging multi-agent pattern: not enforcing too much execution structure and pairing it with dedicated agents for advising and verification. I think it is generally applicable as well. Great read. Here is how it works: Cogentic runs on Gemini and works on open problems in theoretical computer science, starting from the problem statement with no expert hints. The system works in rounds, and the orchestrator decides how many provers to run in each round. Every prover gets one direction to work on, such as a specific bound or a counterexample search, plus a short briefing that a summarizer agent writes from earlier attempts and verifier feedback. Each summarizer writes its briefing independently, so provers in the same round read different summaries of the same history. Each draft goes through two adversarial verifiers. One checks the draft on its own, and the other reads all of the round's drafts side by side to catch shared mistakes. A draft is accepted only if both pass it. The agents share state through two disk documents. A record logs every attempt with the objection it failed on, and a ledger stores verified lemmas and ruled-out directions. An auditor extracts correct lemmas from rejected proofs, verifies them again independently, and adds them to the ledger. A separate process advisor reads the verification logs across rounds and updates the instructions given to provers and verifiers. The orchestrator and the advisor can't give mathematical opinions, so the provers provide all the math. It produced new results on five open problems in online learning, auction theory, and mechanism design, each checked by domain experts. Most problems took around 100 Gemini calls, and the hardest took around 1,000. Paper: https://arxiv.org/abs/2609.40324 Chat with Paper: https://academy.dair.ai/papers/cogentic-multi-agent-orchestration-for-automated-proof-discovery-2609.40324
· 2.7K views
The verifier in Cogentic does not receive the reasoning of the agent searching for a proof. The search agent tries bold ideas, while the verifier rejects invalid proofs. Google Research uses the Gemini-based system to tackle open problems in theoretical computer science without expert hints.
Separate context. In Deep Agents, fork mode passes the parent's conversation history and system prompt to the subagent. By default, isolated mode passes only the delegated task. The latter is suitable for verification with a fresh context.
How to set it up. Install the library with `pip install deepagents`. To run it, you need a model that supports tool calling and a provider API key. Add your own subagent through `create_deep_agent(subagents=[...])`: isolated mode requires `name`, `description` and `system_prompt`, and the agent delegates work using the `task()` tool.
The main agent receives only the subagent's final result. Intermediate search results, file reads and tool calls stay in a separate context. You can set `model` and `tools` separately for the verifier, and the `tools` list completely replaces the inherited one.
In Cogentic, the orchestrator decides how many proof-search agents to launch in each round.
Source
