Every request sent to an AI model uses tokens. These may include system instructions, programme context, uploaded evidence, previous messages, tool descriptions, reasoning, and the final response.

In a small test, inefficient token usage may not matter. In an M&E workflow reviewing hundreds of reports, indicators, interviews, or survey records, repeated context can increase cost and latency.

The objective is not to remove useful context. It is to preserve the information required for accurate and responsible work while reducing unnecessary repetition.

Where do tokens go?

A typical M&E request may contain several different types of context:

Many applications send the same long instructions with every request. In an M&E workflow, repeated material may include:

• The agent’s role and responsibilities.
• Indicator definitions and validation rules.
• Donor reporting requirements.
• Qualitative coding frameworks.
• Confidentiality and data-protection rules.
• Standard output structures.

Prompt caching allows stable prompt prefixes to be reused. Put stable instructions first and place changing user input or programme data later.

Simplified JavaScript example

import OpenAI from "openai";
const openai = new OpenAI();
const instructions = `
You are an M&E indicator quality-review agent.
Check clarity, measurability, disaggregation,
data-source suitability and reporting relevance.
Do not invent baselines, targets or evidence.
`;
const response = await openai.responses.create({
  model: "gpt-5.6",
  prompt_cache_key: "me-indicator-review-v1",
  instructions,
  input: `
Review this indicator:
Percentage of trained participants who apply
the new procedure within three months.
`
});
console.log(response.output_text);

Some M&E tasks develop over several stages. An agent may first review an evaluation matrix, then identify evidence gaps, and finally draft recommendations.

When the objective, assumptions, evidence base, and priorities remain stable, GPT-5.6 can continue from compatible reasoning produced in an earlier response.

Simplified JavaScript example

import OpenAI from "openai";
const openai = new OpenAI();
const first = await openai.responses.create({
  model: "gpt-5.6",
  input: `
Review this evaluation evidence matrix.
Identify weak links, contradictions and missing sources.
`,
  reasoning: {
    context: "current_turn"
  }
});
const second = await openai.responses.create({
  model: "gpt-5.6",
  previous_response_id: first.id,
  input: `
Prioritise the three evidence gaps
that the evaluation team should address.
`,
  reasoning: {
    context: "all_turns"
  }
});
console.log(second.output_text);

Tool-heavy workflows can produce large intermediate outputs. An M&E agent may retrieve hundreds of indicator records, search multiple document collections, compare countries, remove duplicates, and rank evidence.

Programmatic Tool Calling allows GPT-5.6 to compose JavaScript that coordinates eligible tools and processes their results in a hosted environment.

The programmatic stage can filter, sort, join, aggregate, validate, or remove duplicates before returning a smaller result for final interpretation.

Example M&E request

A programmatic stage could:

1. Retrieve records from several tools.
2. Calculate variance between actual and target values.
3. Remove duplicates.
4. Filter records with insufficient evidence.
5. Rank the remaining records by risk.
6. Return only the records needed for final analysis.

Codex may receive guidance from project instructions, AGENTS.md files, skills, tool descriptions, and the current request.

Over time, the same rule may appear in several places. Instructions may also become outdated, conflict with each other, or describe every minor step even when the intended result is clear.

Keep important constraints, but state each rule once and place it in the most appropriate instruction file.

What to audit

Example of a leaner AGENTS.md file

# Project objective
Maintain the M&E reporting application.
# Required standards
- Preserve data-validation rules.
- Do not expose confidential data.
- Use TypeScript with type checking.
- Run tests relevant to changed files.
- Report validation that could not be completed.
# Approval boundaries
Ask before:
- deleting production data;
- changing external services;
- publishing or deploying;
- making purchases;
- expanding the requested scope.
# Definition of done
The requested change is implemented,
type-checked and tested.

A practical implementation sequence

1. Establish a baseline: Record tokens, cost, latency, task success, and human-review findings.
2. Separate stable context: Put reusable instructions before changing inputs.
3. Test persisted reasoning: Use it only for closely related turns.
4. Identify bounded tool stages: Look for filtering, ranking, joining, or deduplication.
5. Audit instructions: Remove duplicated guidance one section at a time.
6. Compare quality: Keep an optimization only when it preserves accuracy, evidence, safety, and usability.

What should you measure?

Token efficiency must not weaken M&E quality

Do not remove indicator definitions, methodological limitations, disaggregation requirements, confidentiality controls, source references, uncertainty, contradictory findings, approval boundaries, or human-review steps merely to reduce token use.

Final optimization checklist

Key takeaway

Efficient token usage comes from careful workflow design: cache stable context, preserve reasoning only when it remains relevant, reduce large tool outputs in bounded processing stages, and keep Codex instructions clear and non-duplicative.

Continue learning

Get 95+ practical AI tutorials for M&E professionals

Explore practical guidance on evaluation design, AI agents, data analysis, reporting, evidence synthesis, data quality, and responsible AI.

Explore EvalCommunity Academy