agentage.hu / Articles / Checklist
Guardrails for News Agents: Step Limits, Budgets and Stop Rules
A practical checklist of limits that separates a fragile demo from a reliable, daily-run newsroom agent system.
Why Unlimited Agents Fail in Production
Unlimited agents create runaway costs and degraded output quality. When a system lacks hard constraints, it repeats the same search queries or rewrites the same summary endlessly. This behavior consumes budget without adding value. The result is a system that works well in a short demo but collapses under real-world load.
Research indicates that token usage explains up to 80% of performance variance in browsing tasks, but only if the task allows it. In sequential workflows, adding more agents often reduces quality by 39% to 70%. The communication overhead fragments the reasoning chain. You lose coherence as the system tries to coordinate too many independent parts.
Coordination is the missing link in most failed deployments. Without a central orchestrator, agents duplicate work. Studies show that uncoordinated systems burn 53% to 86% of tokens on redundant tasks. A single agent with clear boundaries often outperforms a chaotic swarm. You need a structure that directs effort, not just a larger budget.
The distinction between a prototype and a production system lies in these limits. A demo runs once and stops. A production system runs continuously and must self-regulate. You must treat token budgets and iteration counts as first-class primitives. These are not optional settings; they are the foundation of reliability.
Iteration Caps Per Task
Set a maximum number of turns for every specific task. A news monitoring agent should not loop indefinitely trying to find a source. Define a hard cap, such as ten tool calls or five reasoning steps. Once the agent hits this limit, it must stop and report its current state. This prevents infinite loops caused by ambiguous instructions.
Different tasks require different caps based on their complexity. Simple fact-checking might need only three iterations. Complex investigative research could justify ten or fifteen. The key is to define these limits explicitly in the prompt. Do not rely on the model to decide when to stop. It will optimize for completion, not efficiency.
Implement a 'resume' strategy instead of a full restart. If an agent fails at step four, do not reset the entire process. Allow it to continue from the last known good state with a fresh attempt at the failed step. This approach saves tokens and preserves context. It is cheaper and faster than restarting the whole chain.
Monitor the distribution of steps across your workflow. If most tasks hit the cap frequently, your limits are too low or your instructions are unclear. Adjust the configuration based on actual logs. A static rule that never changes will eventually break your system. Use data to refine your thresholds over time.
- Simple fact search: 3-10 tool calls
- Direct comparison: 10-15 calls per agent
- Complex research: 10+ agents with distinct roles
- Hard stop triggers immediate escalation
Cost Ceilings Per Task and Per Day
Establish a token budget for every individual task. Treat tokens as currency with a fixed exchange rate. If a research query exceeds its allocated budget, the system must terminate. This prevents a single complex topic from draining resources meant for the rest of the day. You cannot manage what you do not measure.
Multi-agent systems consume significantly more tokens than single agents. A simple agent uses roughly four times the tokens of a basic chat. A coordinated multi-agent setup can use fifteen times as much. Ensure the value of the output justifies this cost. High-value, parallelizable tasks warrant the expense. Routine checks do not.
Upgrade the model before increasing the budget. Switching to a more capable model often yields better results than doubling the token count on an older one. A smarter model solves problems faster with fewer attempts. Optimize your architecture first. Then scale the budget only if necessary.
Track daily spending against a total ceiling. If the newsroom budget is reached, pause all non-essential agents. Keep critical verification loops running but suspend broad discovery tasks. This ensures the core function remains operational. You avoid a surprise bill at the end of the month while maintaining essential coverage.
Explicit Stop and Escalate Conditions
Define clear conditions that trigger an immediate stop. These include hitting a token limit, reaching an iteration cap, or detecting a confidence drop below a threshold. The agent must not try to 'fix' these errors on its own. It must halt and signal for human review. This is the boundary between automation and chaos.
Escalation paths must be predefined. When an agent stops, it should route the partial result to a human editor. Include the context of why it stopped. Did it run out of tokens? Did it find conflicting sources? Did it hit a safety filter? The editor needs this metadata to make a decision. Blind handoffs waste time.
Use consensus mechanisms for high-stakes decisions. Before publishing a major story generated by agents, require agreement from multiple sub-agents. If the confidence scores diverge, flag the item for manual review. This adds a layer of safety without slowing down routine updates. It balances speed with accuracy.
Implement rollback capabilities for critical actions. Save a snapshot of the state before any irreversible step. If the agent makes a mistake, you can revert to the previous version. This requires immutable storage and clear versioning. It turns a potential disaster into a recoverable error.
Keeping Partial Results When a Limit Hits
Never discard work just because a limit was reached. An agent that stops after eight steps still has valuable data from the first seven. Store these intermediate artifacts in an external database. Reference them later instead of regenerating them. This saves tokens and preserves the history of the investigation.
Use an artifact pattern where agents write to files and pass references. The orchestrator receives a pointer to the result, not the full text. This keeps the context window clean and manageable. It also allows other agents to read the same data without duplication. Structure your output for reuse.
Ensure every artifact includes provenance details. Record who created it, when, and from which sources. Add a brief summary of the evidence found. Without this metadata, the reference is useless. The orchestrator cannot verify the quality of the data it retrieves. Trust requires transparency.
Design your system to handle incomplete chains gracefully. If a task fails mid-way, save the state and alert the user. Do not pretend the job is done. A partial report is better than a hallucinated one. Your editors can finish the work with the available data. Honesty about limitations builds trust.
Key takeaways
- Enforce hard iteration caps and token budgets to prevent runaway loops and cost explosions.
- Use a central orchestrator to reduce error amplification and eliminate duplicate work.
- Define explicit stop conditions that trigger immediate human escalation for high-risk items.
- Store partial results as versioned artifacts to preserve progress when limits are hit.
- Prioritize model capability over token volume to achieve better efficiency.
Related questions
How many iterations should I allow for a news research agent?
Set limits based on task complexity. Simple fact searches need 3-10 tool calls. Direct comparisons require 10-15 calls. Complex investigations may justify 10+ agents with distinct roles. Always define a hard stop to prevent infinite loops and ensure the system reports its status when limits are reached.
What happens if an agent exceeds its token budget?
The system must terminate the task immediately and save the partial results. Do not allow the agent to continue or retry. Route the incomplete output to a human editor with metadata explaining the stop reason. This prevents unexpected costs and ensures no work is lost due to budget constraints.
Is it better to use one strong agent or multiple weak ones?
It depends on the task structure. For parallelizable work like broad research, multiple agents with an orchestrator improve quality by up to 90%. For sequential tasks like writing or coding, a single agent avoids communication overhead and performs 39-70% better. Match the architecture to the workflow.
AI-assisted: drafted with a language model from the author's research notes, then edited and checked before publication. Author: Zoltán Varga, AI visibility researcher and AI ecosystem architect, Budapest.
