Companies rebuild for agent readers: docs, memory, and the org chart all get an agent layer

Docs are now read mostly by agents

Mintlify says agents made 471 million requests across Mintlify-powered sites in October 2026, and that those sites serve over 20,000 companies. The company also cites its State of Knowledge Report finding that 53% of respondents say their company knowledge lives across five or more surfaces. When the reader is mostly an agent, documentation stops being a website and becomes infrastructure: it has to be machine-readable, current and structured, or the agent will work from stale or scattered information. m:…4331

The company offers customer examples for what that shift is worth. Coinbase achieved a 20x faster documentation update speed after implementing Mintlify and estimated $152,000 in support savings year to date due to AI assistant usage, the company says. Anthropic manages over 3,000 pages of Claude Codex documentation with one writer, Anaconda saved over $200,000 in annual headcount costs by avoiding the hire of a dedicated docs engineer, Decagon automates an estimated 95% of documentation updates, and RV Tech saved roughly 120 engineering hours annually by consolidating documentation onto a single platform, according to the company. m:…4331

Self-updating docs need a harness and a memory

Mintlify says it rebuilt its self-updating documentation architecture from the ground up, splitting the system into a background agent for strategy and a runner agent for execution, with memory stored as a git-backed repository containing JSON logs of agent runs. Structured memory schemas allow for offline evaluation of agent performance, the company says, which is what makes the system improvable rather than merely automatic. m:…4843

The results the company reports are concrete: its internal documentation PR acceptance rate increased from 70% to 90%, and its in-house harness, built using Vercel's AI SDK, performed similarly to Pi but used 50% fewer tokens. Breaking changes are shipped to agents by sending plain-text instructions to the background agent, which is instructed to audit its current setup for regressions afterward. The architecture powers the recently launched Autopilot product, the company says, and Mintlify Base is currently in beta. m:…4843

Stop re-explaining the company on every run

Dosu starts from a cost figure that explains why memory matters: the company says 94% of token expenditure in enterprise agents is consumed by re-reading company context, because internal agents primarily spend time sifting through codebases, docs and Slack. The proposed fix is to treat knowledge as a cache, distilling it in the background and injecting it into sessions rather than relying on tool calls, the company says. Frontier models trust Markdown files and native hooks more than information returned from tool calls, Dosu adds, which is why the cache should live where agents already look. m:…1019

The raw material for that memory, the company says, includes human corrections, hard-won context, backtracking and unproductive tool calls. Dosu's own issue triage agent improved after learning that database tables use singular names, a small fact that illustrates the general point. Scoping information to the correct level, whether user, branch, repo, team or organization, is a critical challenge for agent memory, the company says, and teams should start by centralizing knowledge in Markdown files or Confluence. Dosu also reports using Pydantic Logfire to pull production examples into evaluation datasets, and that switching to System One models for extraction reduced P50 latency by approximately half. m:…1019

Manus reorganizes the company around agents

Manus describes going further, rebuilding how the company itself works around agents. The company says Manus-mono serves as a single entry point for agents containing architecture and business context, uses Git worktrees to facilitate parallel development, and lets agents read logs and configuration files to diagnose production bugs. Engineers now manage 10 to 20 requests simultaneously, the company says, and Manus is shifting from human in the loop to human on the loop for AI-driven tasks, with humans retaining final approval authority over release steps. m:…6073

The company is candid about what did not work: autonomous agent loops failed at Manus when applied to product decisions due to lack of testable success criteria. What did work, it says, includes internal agents investigating bug reports and tracing them to specific code changes, code review bots verifying changes with extra rules for sensitive areas like payments, and product managers using agents to benchmark models and build UI demos, with approximately half of product managers shipping features independently. Marketing agents manage their own queues and learn from forwarded Stack context, the company says, and Manus tracks AI spend to evaluate ROI rather than individual performance. The open question is whether human on the loop scales oversight as fast as it scales output. m:…6073