Part 9 of 13 · 4 min read
A fresh context window is not a fresh agent, if the files carry over
Think of a construction crew that changes shift every eight hours. The new shift did not watch the previous one work. They do not need to. The site itself -- what is built, what is marked unsafe, what the plan says is next -- carries everything they need to pick up.
In plain words: part 3 said a scratchpad only remembers what it flagged at write time. A long-running agent faces the same problem one level up: when its context window fills (module 6) or a task simply outlasts one session, the agent itself gets reset to a blank window, and the only thing that survives is whatever it wrote to the workspace this module builds. A progress file -- a scratchpad written specifically so that a different, memoryless instance of the same agent can resume from it -- is what makes that reset survivable. The same idea, applied to standing instructions rather than in-task progress, is what AGENTS.md or CLAUDE.md are: a file read at the start of every session so that project-specific conventions do not have to be re-discovered, or re-explained, each time the window resets. Put together, this is the sense in which an agent is its model plus its harness: the model contributes reasoning that resets every session; the harness, including these files, is what accumulates.
A worked example. Take the mini-workspace scratchpad from part 3 and change what it is written for. A notes-only agent writes step 4: found reference code X77 because that looked worth flagging for answering a later question in the same session. A handoff-style progress file instead writes done: steps 1-4 extracted and filed under refs/; next: step 5 needs the Q3 ledger, not yet fetched; blocked on: none, because it is written for a successor that remembers nothing else at all. The file's content barely changed; its audience did. A workspace-confined agent that reads its own AGENTS.md at session start, then updates a progress.md at the end of each step, can be killed and restarted from either file with no loss beyond whatever happened inside the one step that was interrupted.
About this example About this part
This module's mini-workspace code was built for offloading tool output (part 1), not for session handoff, and has no simulated restart or multi-session experiment; nothing here is measured. "Agent equals model plus harness" and the shift-change framing for long-running work are described in several 2026 practitioner write-ups (Anthropic's own engineering team among them); AGENTS.md itself was proposed as a shared, tool-agnostic convention and placed under the Agentic AI Foundation on 9 December 2025. Last checked: 15 November 2026.
Harness, progress file, AGENTS.md
The harness is everything around the model that persists across a reset: the workspace, the files it reads and writes, the tools it is given, the orchestration code that calls it. A progress file is a scratchpad written for a successor with no other memory of the task, not for the same session later -- so it states what is done, what is next, and what is blocked, rather than whatever looked notable in the moment. AGENTS.md (or a tool's own name for the same idea, such as CLAUDE.md) is a project-root file read at the start of every session, carrying conventions and constraints so they survive a context reset without being re-explained.
Example: the shift-change crew reading the site instead of a handover meeting is a progress file doing the same job: the state that matters is written down where the next shift, or the next session, will look for it on its own.
For the curious Why this is not just module 3's compaction again
Module 3 keeps a long session going by shrinking its own history. A progress file does something narrower and more durable: it assumes the history is simply gone, and asks what the smallest file is that lets someone -- or something -- with zero memory of this session pick the task back up correctly. A good progress file stays short for a different reason than a summary does: not because space is scarce in the next window, but because every sentence in it is a claim the next instance will act on without being able to check it against anything else.