How to Keep Long-Running AI Agents on Track

TL;DR

Keep a long-running AI agent aligned by maintaining a short, authoritative record of the current goal, decisions, progress, limitations, and next action. Update that state whenever evidence changes the assignment, and make every new session read it before working, so stale instructions and noisy history do not control future actions.

Transcript

Three engineers at OpenAI shipped an internal product in about a tenth of the time it would have taken by hand. Roughly 1,500 poll requests. It was a codebase over a million lines by the time they were done. Not a line of it was typed by humans. A few days ago, I made a videos arguing that you should effectively keep your agents desk neat and tidy ... Read More

Key Insights

• Long-running agents amplify both good and outdated judgment because they can continue executing a chosen direction for hours. A slightly incorrect assumption can therefore spread through extensive research, code, analysis, or drafting before a person notices that the work has moved off course.
• A giant instruction manual becomes unreliable when old rules remain beside newer decisions. OpenAI addressed this problem with a short map that directed Codex toward active plans, decision logs, design documents, architecture information, and quality grades relevant to the current work.
• Progressive context shaping is the practice of updating the small set of instructions and decisions that should govern the agent's next actions. The maintained current state receives priority over the accumulated history, allowing evidence from ongoing work to reshape the assignment.
• A strong opening prompt defines the desired result, available materials, permissions, boundaries, quality expectations, and a sensible first action. It remains necessary, but it cannot anticipate discoveries such as weakened research assumptions, architectural dependencies, different customer needs, or low-value early results.
• Anthropic's progress file functions as portable memory between sessions by recording current status, completed work, known limitations, and failed approaches with their reasons. A fresh Claude session can read the file, continue with the next task, and avoid repeating previously discovered dead ends.
• Arize improved a stalled agent by storing its current plan outside the conversation and rebuilding a short plan message for every model call. The agent could then distinguish completed work, work in progress, and the next task despite noisy tool output and intermediate results.
• The storage format is less important than ensuring every next action reads the updated state. The speaker has used a current Markdown file, a plan, JSON, an issue tracker, and carefully revised conversations to carry authoritative decisions across tools and sessions.
• A decision belongs in persistent project state when a fresh agent would cause frustration by missing it. Details that merely explain how the project reached its current position can remain in the historical transcript instead of competing with active instructions.
• More videos from this channel:
  • How will GPT-5 reshape AI and enterprise use
  • Why Do AI Agents Create More Work for Humans?
  • How to roll out AI with team buy-in and minimal resistance
• More from this channel: AI News & Strategy Daily | Nate B Jones

Questions & Answers

Q: How do you keep a long-running AI agent on track?

Maintain a short, authoritative project-state record that tells the agent what is true now, what has been completed, what is in progress, and what should happen next. Update it after consequential decisions or discoveries, then require each new session or action to read it. This prevents stale opening instructions and accumulated tool output from governing later work.

Q: What is progressive context shaping for AI agents?

Progressive context shaping is the practice of revising the small set of instructions and decisions that govern an agent's next actions as the project produces new evidence. It begins with a clear brief, but the current state is updated when research, implementation results, customer evidence, limitations, or failed approaches change the assignment or its definition of done.

Q: Why is one large AI agent instruction file ineffective?

One large instruction file can crowd out the immediate task and become a graveyard of stale rules. Older assumptions remain visible alongside newer decisions, making it harder for the agent to identify which direction has authority. A short map to maintained plans, decisions, designs, architecture information, and quality criteria gives the agent more reliable guidance.

Q: What should an opening prompt for a long agent run include?

An opening prompt should define the desired result, identify the materials the agent may use, establish what it may and may not do, and provide a sensible first move. It should also state known quality standards and identify actions requiring approval, including publication, spending, deletion, or contacting another person. These guardrails let useful work begin safely.

Q: What should an AI agent progress file contain?

A progress file should record the project's current state, completed work, known limitations, failed approaches and why they failed, plus the next task. It may also preserve consequential changes to the goal, constraints, or definition of done. This portable memory helps a fresh session continue effectively without reconstructing the project from a long conversation transcript.

Q: How can you redirect an AI agent without restarting the project?

Change the maintained current-state record so it reflects the revised assignment, then ensure the agent reads that state before its next action. The correction should specify the updated goal, relevant decision, current progress, and next step. Because unfinished work follows the latest authoritative state, the project can change direction without discarding completed work or restarting from the beginning.

Q: Why should current context be separated from conversation history?

Conversation history contains useful evidence, but it also accumulates tool output, intermediate results, obsolete instructions, and the agent's own activity. Separating current context gives the maintained plan greater authority than that noisy transcript. The agent receives a direct answer to what it should do now while the history remains available for supporting detail when needed.

Q: When should a decision be saved in persistent agent context?

Save a decision when it changes the target, definition of done, constraints, permissions, quality expectations, or future actions, especially if a fresh agent session would make a meaningful mistake by missing it. If information only explains how the team reached the current position, it can remain in history rather than occupying the compact, authoritative project-state record.

Summary & Key Takeaways

• Long agent runs need more than a strong opening prompt because projects generate new information as work proceeds. Progressive context shaping keeps a compact statement of the current assignment updated, allowing the agent to follow revised goals, constraints, decisions, and quality standards instead of treating outdated initial assumptions as permanent instructions.

• OpenAI used a short map pointing to active execution plans, decision logs, design documents, architecture information, and quality grades. Anthropic used progress files, structured handoffs, and version history. Although their implementations differed, both approaches gave new agent sessions portable access to the latest project state and known limitations.

• A maintained current-state file separates authoritative direction from the historical transcript. It should preserve consequential changes affecting the target, definition of done, constraints, or next action, while leaving explanatory details in history. The storage format can be Markdown, JSON, an issue tracker, a project board, or an updated conversation brief.