Agent context window full and dropping information? Fixing context overflow

Early in a run the agent is sharp. Twenty steps later it starts forgetting instructions, hallucinating fields, or calling a tool with the wrong arguments — not because the model got worse, but because the context window filled and quietly pushed the things it needed out of scope. Context overflow is a slow degrade, not a crash, which is what makes it sneaky.

What gets lost

Context is finite, and a long run keeps adding to it — every tool result, every reasoning step. When it fills, the oldest content falls out of scope first, and that's often the content the agent most needs:

• Tool definitions get pushed out — so the model calls a tool from memory of its shape, and hallucinates the arguments.
• The original goal drifts out — so the agent optimises for the last thing it read, not the task.
• Earlier results vanish — so it re-fetches what it already had, or contradicts a fact it established ten steps ago.

The tell is that quality falls off as the run gets longer. If your agent is great for the first N steps and unreliable after, suspect context overflow before you suspect the model.

Catch it before it overflows

The fix is to watch context usage mid-run and act before the limit, not after. budget_guard includes a context check and warns as you approach the ceiling — free, no LLM:

budget_guard({ context_pct: 0.88, tokens_used: 88000, token_budget: 100000 })

// → { warnings: [{type:"context_limit", severity:"warning",
//                 suggestions:["summarise old turns", "pin tool defs + goal", ...] }] }

The fix

1. Summarise or prune old turns when context approaches the limit — keep the gist, drop the raw transcript.
2. Pin what matters. Keep tool definitions and the original goal in context at decision points so they never fall out of scope.
3. Store large outputs by reference. A big tool result shouldn't be inlined into context — keep a pointer and fetch on demand.
4. Watch it every step. budget_guard is cheap enough to call continuously, so you act on a warning instead of discovering the overflow in a degraded output.

Diagnose a degraded run

If a run went sideways and you suspect context, send the trace — Snapback returns failure_class: "context_overflow" with the root cause, so you can confirm it before restructuring your context strategy.