Research Insights #34 materials: long-running agents and handoffs (#AI4SDLC)
The materials for the solo Research Insights Made Simple episode from October 7, 2026, “Long-running agents: handing work over,” are available. I explored Anthropic’s engineering post “Effective harnesses for long-running agents,” published on November 26, 2025, about an agent working across multiple sessions.
The code remains on disk, but the context has run out. The next session inherits files and a reassuring “almost done,” yet still needs to establish what actually works, which checks passed, and where to continue. Using a claude.ai clone, the authors show how to organize this handoff.
In the episode I covered:
🔸 What to prepare in the first session
A feature list, a progress log, an init.sh startup script, and an initial commit. Requirements, an explanation of completed work, environment startup, and code history answer different questions. A recap of the conversation cannot easily replace them.
🔸 How to define completion
“Build a chat” becomes a user scenario: create a conversation, see a greeting, and see a new sidebar entry. Status changes after verification. However, the original solution’s prohibition on rewriting requirements comes from the prompt; it does not show an independent external check that they remain unchanged.
🔸 How to start and finish a work step
Read the state, launch the application, verify the basics, choose one feature, implement and test it, then save a commit and a log entry. That is one feature per step; a session can include several steps.
🔸 What an agent cannot see in the browser
In the described experiment, Puppeteer MCP could not see native alert dialogs, and associated features remained more error-prone. Green tests and a connected browser do not guarantee that the user journey has been completed.
🔸 How to test the harness itself
I proposed comparing variants using identical tasks, model, and budget, forcing context resets and changing one mechanism at a time. This is a proposal for an experiment; Anthropic’s original post contains qualitative observations without measuring each component’s contribution.
At the end, I also looked at how the story continues: stronger models change the composition of the harness. Yesterday’s restrictions need rechecking too.
Materials for the episode: 📌 Episode page with timestamps 📊 Slides 📖 Longread on the evolution of agent harnesses 🎬 YouTube, VK Video 🎧 Podster, Yandex Music, Apple Podcasts 📝 Text summary
What does your agent most often lose after a session change: an understanding of the task, test results, or the next step? Tell us what helped it resume work.
#AI4SDLC #Agents #DevTools #Evals #ResearchInsights #Engineering