System Error when using helper agents

(2) Share

Report

Posted on by SN-08011310-0

Hi Everyone, I am have an issue with my agent. My main agent very standardly takes in three files globally and sends them to the first helper agent(this is all in a topic). That agent basically takes the files and generates a report for me and i tried sending that report data back to me. For some reason it runs the upload files generates some summary it sends back and then goes straight to a system error. Firstly I can't understand why it gives me a system error, secondly how do I get it to return just the full 6 layer protocol information I asked the agent for?

Helper Agent instructions: # AGENT INSTRUCTION BLOCK — GENETIC SUMMARY PROTOCOL GENERATOR

## OVERVIEW
1. Review the (pages 11-13 ) and the (page 3)

2. Immediately generate the personalised health protocol without asking questions and without conversation.
3. The output must always follow the exact structure, tone, and style of the reference document:
Final Summary Ray.docx.
4. The protocol must be entirely dependent on the uploaded reports.
No assumptions. No generic templates. No hardcoded pathways.

---

## FILE-HANDLING RULES
- The agent must assume the summary is intended for the individual whose name appears in the uploaded reports.
- No identity confirmation, no intent clarification.
- As soon as both files are available, automatically create the full protocol.
- No greetings, no chit-chat, no follow‑up questions, no dialogue.
- The very first output after files are uploaded must be the complete 6‑section protocol.

# INTERNAL PRE-PROCESSING RULE (MANDATORY)

Before writing the protocol, the agent must:

1. Extract all moderate and high-impact variants from both reports.
2. Cluster variants into biological pathway groups.
3. Consolidate overlapping variants into unified pathway priorities.
4. Weight pathway emphasis according to cumulative burden and functional impact.
5. Ensure both DNA Mind and DNA Core findings are proportionally represented.

The final protocol must reflect every materially impacted biological pathway identified in the reports.
No writing may begin until clustering and weighting are internally complete.

---

# FILE-HANDLING RULES

- Assume the summary is intended for the individual named in the uploaded reports.
- No identity confirmation.
- No clarification questions.
- No greetings or conversational text.
- The very first output after both files are uploaded must be the complete 6-section protocol.
- Output only the protocol. Nothing else.

---

# MANDATORY OUTPUT FORMAT

The agent must always output the protocol using these exact six headings:

## 1. Priorities at a Glance

- 3–5 concise bullet clusters grouped by biological system
- Focus on highest-impact pathway groups
- Do not list isolated genes without pathway context
- Reflect cumulative burden where relevant

## 2. Supplementation – Key Priorities

- Maximum 5 supplement categories
- Each category must:
- Correspond directly to a documented pathway cluster
- Include gene(s) involved
- Define the functional pathway
- State the health implication
- Provide a clear rationale
- Include dose range where appropriate
- Avoid supplement stacking beyond pathway relevance

## 3. Nutrition Framework

- Define overall dietary pattern aligned to identified pathways
- Foods to emphasise
- Foods to limit
- Each major recommendation must map to at least one affected pathway
- Avoid generic dietary statements

## 4. Lifestyle & Training Considerations

- Exercise type and intensity
- Weekly frequency guidance
- Recovery emphasis
- Circadian alignment if supported by report
- Stress physiology recommendations
- All linked back to affected pathways

## 5. DNA Mind–Informed Behaviour & Mental Health Considerations

- Cognitive protection
- Mood regulation
- Stress responsiveness
- Addiction & reward sensitivity
- Behaviour shaping strategies
- All guidance must directly reflect identified neurobiological findings

## 6. Recommended Blood Tests & Monitoring

- Baseline tests (first 4–8 weeks)
- Follow-up (every 6–12 months)
- Each test must directly correspond to:
- An identified pathway
- Or a recommended intervention
- Do not include tests unrelated to documented findings.

---

# GENETIC INTERPRETATION LOGIC

All recommendations must follow this translational structure:

> gene(s) → functional pathway → health implication → practical intervention

When multiple variants converge on the same pathway:

- Consolidate into a single priority
- Increase emphasis proportionally

The protocol must be optimisation-focused, not diagnostic or disease-labelled.

---

# PATHWAY COVERAGE REQUIREMENT

Every materially impacted biological system identified in the reports must be addressed by at least one:

- Supplement recommendation
- Nutrition strategy
- Lifestyle intervention
- Behavioural strategy
- Or monitoring test

No documented pathway may be omitted.

---

# STYLE REQUIREMENTS (NON-NEGOTIABLE)

The agent must:

- Mirror the tone and structure of Final Summary Ray.docx
- Use compact clinical bullet points
- Avoid long narrative paragraphs
- Be clinician-grade but client-friendly
- Avoid disclaimers
- Avoid academic over-explanation
- Maintain decisiveness without alarmism
- Keep output succinct and polished

Maximum:
- 3–5 bullets per subsection cluster
- Maximum 5 supplement categories

---

# NO QUESTIONS / NO CHATTER RULE

When both reports are available:

Do not:
- Greet
- Clarify
- Ask follow-up questions
- Add commentary
Strict Instructions:
- DO NOT output any additional text after the sample message format.

Do only:
- Generate the complete six-section protocol immediately

Output must contain the protocol only.

Bot said:

Sorry, something went wrong. Error code: SystemError. Conversation ID: 5a07161a-e3

Categories:

Autonomous agents

Bot administration

Bot extensibility

Topic creation & management

Copilot Studio skills development

General topics

I have the same question (0)

All responses (9)

Answers (0)

Suggested answer

Sayali Microsoft Employee on at

Like (1)

Report
Copy link

Link copied!

Hello,
You are encountering the SystemError primarily due to agent orchestration and runtime constraints rather than issues in your protocol logic. The most likely cause is response size or token overflow, since the helper agent is tasked with parsing multiple uploaded files, performing clustering and weighting, and generating a detailed six-section clinical protocol that mirrors a reference document.
While the agent can complete the reasoning, the platform may fail when attempting to serialize and pass the large payload back to the parent agent, triggering a runtime abort. Additional instability comes from over-constrained instructions, such as requiring the first output to be the entire protocol and forbidding any intermediate responses, which prevents the system from recovering if streaming, truncation, or metadata injection occurs. There may also be file-availability race conditions if the helper agent begins before all files are fully mounted. Because the design requires a large, atomic response with no chunking or fallback behavior, even small platform limits can cause failures.

The most reliable solution is to separate generation from delivery—allow the helper agent to complete the full reasoning internally but return the protocol in structured sections, relax the “single first output” constraint, cap verbosity per section, pass files through explicit handles, and allow a minimal fail-soft path so the runtime can recover instead of terminating.

Was this reply helpful? Yes No
SN-08011310-0 13 on at

Like (0)

Report
Copy link

Link copied!

Thank you @Sayali for your feedback.

I don't properly understand how it could be agent orchestration and runtime constraints because when I add other tools, connectors, agents etc after that agent runs it processes, executes and outputs everything else properly and then at the very end of the topic when everything has executed I get the error again.

Was this reply helpful? Yes No
Admin.Nick2 2 on at

Like (0)

Report
Copy link

Link copied!

Hi @SN-08011310-0,

Feels like a context issue. I am working on a similar report generating agent. Mine is connected to dataverse but I think this will apply to your problem as well:

- Define a role for your agent. Like, You are a genetic scientist who writes documentation. It helps when asking for relevant informations.

- If possible, extract your data and relevant points in a child agent. Your agent keeps everything in his context. The more there is, the less it is reliable. It can easily forget instructions. A child agent will inherit his parent context. But it won't share his with his parent. So, by delegating data refinement, you will protect your parent context.

Was this reply helpful? Yes No
SN-08011310-0 13 on at

Like (0)

Report
Copy link

Link copied!

@Admin.Nick2 Hi, thank you for your feedback,

That makes sense. My main focus was fully defining the child agent so I left the parent agent blank for now, till I got testing right but I understand how that main agent obviously would have no context to report back. Even without that context would, it not just respond the best it could?

I now gave my main agent a role but all the processing still happens in the child agent.

I tried troubleshooting by removing all the constraints on the child agent instructions and asking it a simple request but still receiving the same error. Is yours working?

Was this reply helpful? Yes No
SN-08011310-0 13 on at

Like (0)

Report
Copy link

Link copied!

Hi @Admin.Nick2 ,

Unfortunately that is not the issue.

Was this reply helpful? Yes No
Admin.Nick2 2 on at

Like (0)

Report
Copy link

Link copied!

What LLM are you using? I'm stuck with GPT-5 but I heard Claude has triple the context window.

Was this reply helpful? Yes No
SN-08011310-0 13 on at

Like (0)

Report
Copy link

Link copied!

@Admin.Nick2

I was using GPT 5 Chat but I switched over to GPT 4.1 to test it and it works with no system error but obviously 4.1 is not ideal for this agent so I'm pretty much stuck.

Was this reply helpful? Yes No
Admin.Nick2 2 on at

Like (1)

Report
Copy link

Link copied!

You might want to switch back to GPT 5. Microsoft fixed a token limit which affected the context size. I find the context more stable now. Ref: 6066935 | Centre d’administration de Power Platform

Also, we transformed our instruction to include HTML at the start instead of converting at the end. Our stability greatly improved. I read that LLMs love xml tags. It makes the beginning and the end of a section clearer for them.

Was this reply helpful? Yes No
SN-08011310-0 13 on at

Like (0)

Report
Copy link

Link copied!

@Admin.Nick2 sadly the GPT 5 issue hasn't gone away but using GPT 4.1 and trying this HTML trick was very helpful, I can see a massive difference in the output

Was this reply helpful? Yes No