Agent settings
Agent-side knobs that apply instance-wide — how many messages an agent keeps in context, how many tools it can call per turn, how issues find their agent when nobody's assigned, and whether agents learn from past successful runs.
What this page is for
Every agent has its own personality, tool grants, and model. This page is for the things that are not per-agent — the deployment-wide defaults that shape how every agent runs. Four cards: Behavior, Compaction, Learning, and the prominent How issues find their agent dispatch strategy card.
Behavior
| Field | What it controls | Default |
|---|---|---|
| Max History Messages | How many prior turns are included in the agent’s context on each call | 20 |
| Max Tool Calls Per Turn | Safety limit on how many tools an agent can invoke in one turn | 25 |
History messages is the rolling window. At 20, the agent sees its last 20 messages plus the system prompt. Higher values give more context at higher token cost; lower values force the agent to rely more on memory and the data store. Compaction (below) automatically summarizes older messages when the window fills.
Tool calls per turn is a runaway-loop guard. An agent that calls 25 tools in one turn is almost certainly stuck. The cap aborts the turn and surfaces an error. If you have a legitimate tool-heavy workload that hits this limit, raise it — but investigate first; 25 is already high.
Compaction
Compaction summarizes old turns into a single system message when the context window fills up. Without compaction, conversations that grow past the model’s context hard-stop with an error.
| Field | Notes | Default |
|---|---|---|
| Enable automatic compaction | Master switch | On |
| Threshold (% of context window) | When to fire compaction | 80 |
| Keep Tail Messages | How many recent messages remain verbatim after compaction | 6 |
At 80% of the model’s context limit, the older half of the conversation gets compressed into a summary message. The last 6 messages remain intact so the agent preserves recent flow. The agent sees a shorter, still-coherent conversation and continues.
Turn compaction off if your workload is short conversations only — you’ll save a bit of per-turn latency. For long-running agent sessions, leave it on or they’ll hit the context wall eventually.
Learning
The learning system extracts reusable lessons from successful agent runs and applies them to future work. It’s off by default; opt in if you want agents to improve over time.
| Toggle | When on |
|---|---|
| Enable learning system | Master switch; off = no extraction and no application |
| Require human approval before applying new learnings | Learnings land in Review queues → Learning review before they go live |
| Allow agents to propose new learnings | Off = only admin manually-entered learnings are used |
| Field | Notes | Default |
|---|---|---|
| Max Extractions per Agent per Day | Daily budget for automatic extraction calls | 10 |
| Default TTL (days) | How long approved learnings stay active (0 = forever) | 90 |
The Advanced tuning disclosure hides six additional knobs: minimum confidence before a learning is retained, cap on how many learnings are injected per turn, reinforcement threshold before confidence auto-boosts, max pending queue size per agent, minimum tool calls required before extraction fires, and a daily-budget fraction above which extraction skips entirely. Defaults are chosen to be reasonable — leave them alone unless you understand the tradeoff.
When to enable learning:
- Long-running deployments where the same kinds of tasks recur
- You can invest a few minutes per week in the learning review queue
When to keep it off:
- Short-term or single-purpose deployments
- You don’t want the review burden and prefer to tune personalities directly
Auto-dispatch — how issues find their agent
When an issue is created without an assignee, the heartbeat sweep picks one. Three strategies:
| Strategy | Cost per dispatch | How it works | Best for |
|---|---|---|---|
| PM-Delegated (default) | $0 — uses existing PM context | Notify the PM agent and let it pick the assignee on its next heartbeat | Default for most teams — the PM sees the full project context |
| Keyword | $0 — pure text matching | Match issue text against agent descriptions by token overlap | Fast, deterministic, no LLM cost; good for stable teams |
| Bidding | ~$0.002 per dispatch | Every specialist gets a bid LLM call. Highest confidence wins | Best matching quality, at small per-issue cost |
PM-Delegated is the default because the Project Manager has the project’s goal context and can make nuanced assignments — the keyword approach can’t distinguish “write about pricing research” from “investigate pricing for research.” The PM sees the whole picture.
Switching to Keyword removes the LLM dependency and is faster, but it only compares words. If your agent descriptions are well-written, it works surprisingly well; if they overlap, assignments can be arbitrary.
Bidding is the most accurate for diverse teams — each specialist gets a small bid call and self-scores its fit. Best when you have 10+ specialists and issue routing matters. The minimum confidence slider sets a floor; bids below it are discarded. Fall back to keyword uses the keyword strategy if no bid clears the floor — prevents unassigned issues when confidence is universally low.
Common pitfalls
Cranking Max History to reduce hallucination. More history usually doesn’t help — it inflates cost and can hurt attention. If an agent is forgetting things, the fix is better memory / data store usage, not bigger history.
Turning off compaction “for simplicity.” Long-running sessions will hit the context wall without compaction. It’s not optional once you’re running real workloads.
Enabling learning without committing to the review. Pending learnings pile up. The extraction system quietly consumes budget on every successful run. Either commit to reviewing weekly or turn it off.
Picking Bidding for a small team. If you have three specialists, Bidding is overkill — the LLM cost adds up and the matching quality gain is negligible. Stick with PM-Delegated or Keyword.
Where to go next
- Agents concept — what specialists are and how they work
- Review queues — where approved learnings and pending learnings live
- Budget — cost context for per-dispatch LLM calls