The agent infrastructure layer is stabilizing fast. This week brought forked context windows, managed credential systems, and a new flagship model built for autonomous work—plus a glimpse at what happens when you run 10,000 agents for 88 hours straight.
Organizing Context in a Multi-Agent Harness
LangChain's deepagents framework now supports forked subagents that inherit the supervisor's full conversation history instead of starting fresh. This matters because it eliminates redundant work like re-reading files or re-fetching data, and it pairs well with prompt caching to cut token costs when subagents need to continue rather than verify. The framework exposes two modes: isolated for independent verification tasks, and fork for continuation work where context inheritance reduces both latency and spend.
const subagent = await supervisor.fork({
mode: 'fork', // inherits full conversation history
task: 'refactor the auth module based on the design we discussed',
cachingEnabled: true
});OpenAI reports Navier-Stokes singularity find in 88 hours using Astra-next, roughly 10,000 agents and 130B tokens (>$40M)
OpenAI announced that a swarm of roughly 10,000 agents consumed 130 billion tokens over 88 hours to produce a claimed solution to a Millennium Prize Problem, at a cost exceeding $40M. The system used multi-agent reinforcement learning trained over a year, with agents autonomously organizing collaboration at runtime. Mathematical validation is still pending, but the announcement signals that test-time compute at massive scale is now operationally feasible—and extremely expensive. If you're running agent fleets in CI or on customer behalf, this is a preview of billing runaway at production scale.
Connections: Managed credentials and per-caller identity for Managed Deep Agents
LangSmith's Connections system separates credentials from code and resolves user-owned OAuth tokens at runtime, so agents can act on behalf of specific users without hardcoding secrets. The platform handles OAuth flows automatically and stores credentials in the workspace, enabling rotation without redeployment. This is critical for production agents that need to respect per-user permissions when calling GitHub, Slack, or internal APIs.
# Agent config with per-caller identity
connections:
github:
type: oauth
scopes: [repo, workflow]
slack:
type: oauth
scopes: [chat:write, channels:read]OpenAI Releases GPT-6 Astra for Coding and Computer Use
GPT-6 Astra ships with a one-million-token context window, 72.6% on OSWorld 2.0, and a hallucination rate of 4.2% (down from 12.2% in GPT-5.6 Sol). It's designed for multi-step task execution and can interact with graphical interfaces, but it's also the first OpenAI model classified at critical cybersecurity capability level after discovering vulnerabilities during testing. For teams building agentic CI/CD workflows, the expanded context window and lower hallucination rate should improve reliability, but the security classification adds compliance considerations for production deployments.
These updates show the agent layer maturing in parallel: better context management reduces token waste, managed credentials enable per-user agent actions, and new models offer more headroom—but the OpenAI swarm experiment is a reminder that cost observability and budget guardrails aren't optional anymore. 📊