Conversation as Collaboration

The interaction between developer and agent is not command-and-response it resembles collaboration between colleagues where both parties contribute to outcomes. Understanding this collaborative dynamic transforms how developers approach agent work.

Session lifecycle

Agent sessions have distinct phases that require different management strategies:

• Initiation Establishes working context. Claude Code loads CLAUDE.md files automatically; Codex scans for AGENTS.md files.
• Active conversation Builds shared understanding through turns. Each exchange adds to context: files read, commands executed, decisions made.
• Degradation The "lost in the middle" effect reduces attention to mid-context information. Long sessions accumulate contradictory instructions.
• Termination Intentional resets preserve momentum by capturing decisions. Forced termination risks losing ephemeral context.

Session commands

Resume the most recent conversation:

claude -c
claude --continue

Resume by session ID:

claude -r abc123
claude --resume abc123
claude --resume  # Interactive session picker

Managing context

The feedback loop

Agent interaction follows an iterative pattern: request → response → evaluation → feedback → refined response.

Effective corrections:

• Include file path and location
• Describe what is wrong AND what should happen
• Provide context for why the change matters
• Keep scope focused to one issue

Ineffective corrections:

• "Fix the bugs" without specifying behaviors
• "Make it better" without criteria
• Multiple unrelated issues bundled together

Context provision strategies

Upfront context accelerates initial responses but risks overwhelming the agent. Works well for stable project knowledge architecture patterns, coding conventions, key file locations.

Autonomous exploration lets the agent navigate using search and read tools. Suits discovery tasks where relevant files are not known in advance. Adds latency but produces targeted context.

Strategy selection

When to reset versus continue

Continue the session when:

• Working on the same feature or related tasks
• Building on established context and prior decisions
• Maintaining momentum on multi-step work
• The context contains accurate, relevant information

Start fresh when:

• Completing a task and moving to unrelated work
• Context window fills with outdated or irrelevant history
• The agent starts "forgetting" instructions or reverting to corrected patterns
• Debugging sessions have accumulated failed approaches
• Context utilization exceeds roughly 80% of practical limits

Session transition technique

Before ending productive sessions, capture value:

1. Ask the agent to summarize key decisions, implementation details, and open questions
2. Copy the summary to a scratchpad file or directly into the new session
3. Clear or exit the session
4. Start fresh with the summary as initial context

A useful prompt for handoffs:

"Write a summary of our session including all decisions made, changes implemented, and open issues. Another engineer will pick this up."

Codex task management

Codex approaches sessions differently through its task-based model. Each task runs in an isolated sandbox preloaded with repository content. Tasks complete independently, producing reviewable results.

The philosophical difference: Claude Code treats sessions as extended collaborations; Codex treats tasks as discrete units of work. Long exploratory sessions favor Claude Code's model; batched independent tasks favor Codex's isolation.