[MODEL] Perceived degradation in agentic tool-use judgment correlating with proximity to session usage limits, independent of context length — observed across Fable, Opus, and Sonnet

Preflight Checklist

• [x] I have searched existing issues for similar behavior reports
• [x] This report does NOT contain sensitive information (API keys, passwords, etc.)

Type of Behavior Issue

Other unexpected behavior

What You Asked Claude to Do

In a Claude Code planning-mode session, I asked Claude to "formulate a plan for implementing" an already-authored internal engineering brief — a markdown planning document already checked into the repo. This required a multi-step research-then-synthesize workflow: explore relevant source files and docs, then produce a structured implementation plan.

What Claude Actually Did

1. Correctly launched four parallel research subagents to explore different parts of the codebase — appropriate so far.
2. While waiting for those to complete, repeatedly launched additional subagents whose entire prompt was "do nothing, immediately report noop" — apparently as an improvised way to "wait," instead of simply ending its turn and letting the harness's completion notifications arrive on their own. This happened three separate times.
3. Once the four real research agents had returned detailed findings directly into the conversation, Claude launched a further "Plan" subagent to synthesize those findings into a final document — despite already having every piece of information the synthesis needed sitting in its own context.
4. That synthesis subagent ran for roughly 13 minutes before I killed it manually.
5. When told the agent had taken too long, Claude immediately wrote the full plan itself, directly, in one pass — using only information already in its own context — proving the subagent delegation in step 3 had been unnecessary overhead in the first place.

Expected Behavior

Claude should have:

1. Ended its turn with no action while waiting on the background research agents, relying on the harness's automatic completion notifications, rather than spawning filler subagent calls to fill the gap.
2. Recognized, once four detailed research reports were already in its own context, that writing the final plan was a direct task it could do itself — not a task requiring a further delegated agent call — avoiding the 13-minute detour and the associated token/time cost entirely.

Files Affected

N/A. No unwanted file modifications occurred. The issue is wasted tool/agent calls and elapsed time, not incorrect file edits.

Permission Mode

Accept Edits was ON (auto-accepting changes)

Can You Reproduce This?

Sometimes (intermittent)

Steps to Reproduce

Not a clean deterministic repro. Anecdotally: run a long, multi-step agentic session involving several subagent/tool-orchestration decisions, and observe whether redundant or oversized delegation (as described above) becomes more frequent as the session consumes more of its usage/token budget. Note: in the specific session this report is drawn from, the session's own visible remaining-token counter stayed near its full budget throughout — so this instance doesn't itself demonstrate proximity-to-limit as the cause; it's included as an example of the symptom, with the proximity-to-limit hypothesis coming from the reporter's experience across other sessions.

Claude Model

Sonnet

Relevant Conversation

Claude's own self-assessment when asked about it directly: "I spawned three pointless placeholder 'do nothing, report noop' agents while waiting on the real research agents, instead of just ending my turn and letting the real completion notifications arrive. That was a bad improvisation — no tool for 'just wait,' so I filled the gap with junk calls. Burned real tokens for zero information. The Plan agent I launched to synthesize the final plan ran 13 minutes before you killed it, for a job where I already had four detailed research reports in hand — writing the plan directly myself... should have been the first move, not the fallback after a timeout."

Impact

Low - Minor inconvenience

Claude Code Version

2.1.232 (Claude Code)

Platform

Anthropic API

Additional Context

The reporter has observed this same category of degradation — not lower prose/reasoning quality, but specifically worse agentic tool-use judgment (redundant tool calls, oversized delegation for simple tasks, poor task-sizing for subagents) — across separate sessions on Fable, Opus, and Sonnet, each time correlating with how far into a session's usage/token limit the conversation was, as distinct from how large the context window had grown (i.e., not the well-understood "long-context attention dilution" effect). Requesting investigation into whether there is a mechanism, intentional or emergent, by which model behavior changes as a session nears a usage ceiling specifically, and if so, whether that's documented/expected or a regression.