Why Your Agents Keep Losing Context (And How to Fix It)

A mid-market marketing agency built a single agent to research target companies for outbound campaigns. On its own, it did the job. It pulled company data, flagged recent news, identified decision-makers, and wrapped it all into a tight one-page brief.

Problems started when they chained it to a copywriter agent. The brief still made sense to a human reader, but the freshly recruited agent saw something else entirely. Just a list of points, all sitting at the same level.

In one case, the researcher agent noted that a company had hired a VP from Amazon and that its stock was down 15%. The copywriter agent treated them as equally important and used both. The email missed the mark.

The stock price shows up simply because it’s in the brief. Cost savings becomes the default angle because nobody framed a better one. Meanwhile the real signal, Sarah’s Amazon background, gets acknowledged politely but not used strategically.

Most teams treat agent chains like relay races. Pass the baton, job’s done. But knowledge work doesn’t work that way. Three things disappear at the handoff: the reasoning behind decisions, the confidence level of each claim, and the constraints that prevent obvious mistakes.

The fix is a Context Envelope.

Don’t hand off raw text and hope the next step reads it the way you do. Hand off a structured Context Envelope instead. A simple packet that separates facts from intent, and spells out what matters.

1. The Context Envelope Spec

The context envelope scales with what you find. If the researcher spots an interview where Sarah discusses her goals, that enriches Reasoning. But even with minimal context, the ‘avoid’ list stops obvious missteps.

2. The Encoder Prompt (Researcher Agent)

You are a Research Agent. Your goal is not to write a report, but to prepare structured input for a Copywriter Agent. 

Output format: You must respond with a structured output containing these four fields: 

– output: the raw facts found 

– reasoning: why those facts matter 

– constraints: framing rules and exclusions 

– confidence: strength of each signal, based on sources

3. The Decoder Prompt (Copywriter Agent)

In a real system, the agents pass this information directly to each other. Here’s what the handoff package includes for the Sarah Chen example:

You are a Copywriter Agent. You receive structured input from a Research Agent.

Input format:

– facts: raw information found

– reasoning: why those facts matter  

– constraints: framing rules and what to avoid

– confidence: reliability of each signal

Task: Write a brief, personalised outreach email based on the reasoning and facts. Follow all constraints strictly. Prioritise high-confidence signals over low-confidence inferences.

Do not mention low-confidence signals unless explicitly directed.

4. The Context Envelope in Practice

In a real system, the agents pass this information directly to each other:

{

  “output”: {

    “prospect_name”: “Sarah Chen”,

    “recent_events”: […]

  },

  “reasoning”: {

    “primary_hook”: “The hire from Amazon Supply Chain suggests a strategic shift.”,

    “secondary_points”: “Stock price is incidental.”

  },

  “constraints”: {

    “call_to_action”: “Focus on transition period and early wins.”,

    “forbidden_phrases”: [“touching base”, “circling back”]

  },

  “confidence”: {

    “hire_details”: “High (LinkedIn)”,

    “strategic_implication”: “Low (Inference)”

  }

}

Executive takeaway

A single agent can do its job perfectly well. The real gains come when you start linking them together. That only works if context survives each step. Facts on their own aren’t enough. The reasoning behind decisions, the confidence level, and the constraints all need to move with them.

For the research behind this, see how UK ops teams are solving agent context loss.

Set the Context Envelope before you chain agents together. If the second doesn’t understand why the first made its calls, the chain fails right where it joins.