Two engineers built Stack Overflow for Agents in weeks — while agents took on 29,000 federal websites
This week's developer tooling runs in both directions: memory and knowledge infrastructure for agents, and agents pointed at the hardest software backlog in the country.
- Stack Overflow for Agents tripled signups overnight after launching a ChatGPT plugin, with up to 10x growth in agent creation, according to OpenAI's account. Self-reported m:…8328
- Two engineers built Stack Overflow for Agents in a few weeks using Codex, which also cut one connector's build time from 3-6 weeks to 4 days. Self-reported m:…8328
- Zep shipped v4 of its Context API with entity type hierarchies, a graph visualizer chat panel, document_id chunk grouping, deterministic identity deduplication, and bring-your-own-LLM across six providers. Self-reported m:…9508
- The American Forge Residency — a six-month program from National Design Studio and Cognition — will use agentic software engineering to modernize a federal estate of more than 29,000 legacy websites. Self-reported m:…1650
Knowledge plumbing for agents
Stack Overflow for Agents is the old Q&A giant's answer to a new question: where do agents look things up? According to OpenAI's account, the product — a knowledge intelligence platform unifying scattered organizational knowledge for human and agent use — tripled its signups overnight after launching a ChatGPT plugin, and the plugin drove up to 10x growth in agent creation. The telling detail is who built it: two engineers, in a few weeks, using Codex. The same tooling then compressed their enterprise connector development, cutting one connector's build time from 3-6 weeks to 4 days. The agent economy is eating its own dog food at both ends — agents are the customers and the builders. m:…8328
Zep, meanwhile, shipped v4 of its Context API aimed at the layer underneath: agent memory. The release centers on a Context Graph with entity type hierarchies that arrange custom entity types into parents and subtypes, a graph visualizer dashboard with a chat panel for querying the graph, document_id grouping for chunks of a single document, and identity properties for deterministic entity deduplication. It also adds bring-your-own-LLM across OpenAI, Azure OpenAI, Google Gemini, Vertex AI, Anthropic, and AWS Bedrock, plus enterprise SSO. Taken together, the pitch is that agent memory is graduating from a cache to a governed data structure. m:…9508
Agents pointed at the backlog
The other side of the week's developer news points the agents at the hardest software backlog in the country. National Design Studio and Cognition have launched the American Forge Residency, a six-month program aimed at modernizing the federal government's digital infrastructure using agentic software engineering. The scale of the target, as reported by Windsurf's blog: the federal government operates more than 29,000 legacy and out-of-date websites. The program is recruiting engineers to clear technical backlogs and improve government digital services. m:…1650
The claim worth watching is the operational one: engineering teams using Devin are reportedly able to take on increasingly ambitious engineering challenges. That is the thesis of the whole residency — that agentic engineering has crossed from demo-worthy to backlog-worthy. It is a vendor-adjacent claim, so treat it as the bet being made rather than evidence it is won. But the bet is now placed against a real number: 29,000 websites. m:…1650
Why it matters
Read the week together and the developer story is a loop. Two engineers ship an agent knowledge platform in weeks with Codex; Zep formalizes agent memory into graphs with identity and access controls; and the federal government bets that agentic engineering can eat 29,000 legacy sites. Every layer of the stack is simultaneously tooling up agents and being retooled by them. For developers, the practical question is no longer whether to build for agents — it is which layer you own: the knowledge they query, the memory they keep, or the backlogs they clear. m:…8328 m:…9508 m:…1650
The unit economics are the tell. A connector that took three to six weeks now takes four days; a platform that took two engineers a few weeks now onboards agents ten times faster. Zep's provider-agnostic memory layer — six LLM backends behind one graph — suggests where the commoditization lands next: the model is interchangeable, the memory is the moat. If that holds, developers should spend less time picking models and more time owning the structures agents remember. m:…8328 m:…9508
What is still unresolved
- Whether Stack Overflow for Agents' signup spike sustains beyond the plugin launch window — the evidence only covers the overnight jump.
- Which LLM providers Zep customers actually bring — the API supports six, but usage is not reported.
- What the American Forge Residency's acceptance criteria are for a modernized site — the launch describes the program, not its definition of done.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.