Agent Mode Issues
When agent runs fail, stop early, hit the wrong tools, or produce confused output.
Agent run hits step ceiling without finishing
Default cap: 20 steps. Run gets cut off with "Max steps reached."
- The goal might be too broad — "research the entire field of X" needs more than 20 steps
- The planner might be looping (calling the same tool repeatedly with similar inputs)
- A tool returned errors causing the planner to retry instead of moving forward
Fixes:
- Refine the goal to be more bounded ("the top 3 papers on X" instead of "everything about X")
- Try a different planner model — reasoning models (o1, GPT-5) are better at planning
- Stop early once you have what you need — partial results are fine
Agent run hits cost ceiling
The cost cap is $0.50 by default. The run halts and synthesizes a partial answer.
To raise the cap: not currently exposed in the UI. The default is conservative for first-time users. If you regularly need more, talk to a real human — we can increase your account's cap.
Run completed but the answer is just JSON dump
The synthesis step (final summarization) sometimes echoes raw tool output instead of paraphrasing it. Known issue tracked as synthesize_now() prompt sharpening. Workarounds:
- Re-run with a more specific goal — gives the planner less room to amass raw data
- Pick a model with better instruction-following for the planner (choose a catalog model with tool support, such as the configured Pro planner)
- Stop the run earlier (when the timeline shows you've gathered enough), and ask the model to summarize in a follow-up chat message
Wrong tool was picked
The planner chose web_search when you wanted quote_lookup (or vice versa). Reasons:
- The goal phrasing didn't include strong tool signals — "What's the price of AAPL?" prefers
quote_lookup; "What's happening with Apple stock?" might preferweb_search - Some models default to web_search for almost everything
Fix: be explicit about the tool when starting the run — "use quote_lookup to get the current AAPL price" — or pick a different planner model.
The 6 tools and their best-fit cases are documented in The 6 Tools.
Stop button does nothing / 404s
Two scenarios:
- The run finished naturally before stop landed. The stop endpoint returns 404 because the registry already cleaned up. The UI should show the completed answer; the 404 isn't an error.
- Network blip in flight. Try again — the stop endpoint is idempotent.
If the timeline keeps streaming but stop has been clicked: the soft-stop is honored at the next iteration boundary, not mid-tool-call. So if the agent is in the middle of a slow web fetch, stopping takes a few seconds to take effect.
Timeline shows ghost steps or duplicates
Ghost steps usually come from step_cost-only events (no tool invocation). Should be filtered out by the reducer; if you see them, the bundle might be stale. Hard-refresh the page (Cmd/Ctrl+Shift+R).
If duplicates persist after refresh, talk to a real human — could be a bug in the SSE reducer.
"Cost shows $0.00 but tokens were used"
Known issue: agent_runs.total_cost_usd is always 0 because the dispatcher doesn't yet expose pricing lookup. Tokens are correct, just the dollar conversion is missing. Tracked for fix; not blocking.
To estimate cost: total_input_tokens + total_output_tokens × the model's per-million rate (visible in the model picker).
Agent answer doesn't appear in chat history
Agent runs append a synthetic user goal + assistant answer to the chat history when they complete. If the answer isn't appearing:
- The run might have failed mid-synthesis — check the timeline for an error step
- You might have switched conversations during the run — the answer goes to the conversation that was active when the run started
- Hard-refresh the page
Recent runs are also visible in the icon-rail History panel.
Agent ignored a tool I wanted
Some tools are L3 (external writes) and aren't shipped yet — send_email, post_to_slack, create_calendar_event, etc. The agent currently has 6 read-only tools. If you ask for something requiring an L3 tool, the agent will say so and fall back to a regular text answer.
Roadmap: L3 tools will land with permission gates and audit logging. No ETA.
Where to next
- What is Agent Mode? — overview
- The 6 Tools — tool reference
- Examples — what works and what doesn't
- Costs and Limits — full caps reference
- Stopping & Reading Timeline — soft-stop details