Why Agent Settings Matter for Backtesting Speed
If you've ever watched an MT5 backtest crawl through years of tick data, you already know the frustration. The default agent configuration leaves performance on the table — often a lot of it. Most traders never touch the agent settings, running single-core tests when their machine has four, eight, or even sixteen cores sitting idle.
MT5's Strategy Tester uses something called local agents. Each agent is essentially a separate process that can run a backtest pass independently. When you enable multiple agents, the tester splits the work across your CPU cores. The result? A backtest that took 20 minutes on a single core might finish in under 3 minutes on a properly configured 8-core machine. That's not marketing hype — it's how the architecture works.
This guide walks you through exactly how to configure those agents, what settings actually matter, and where people commonly mess it up. You'll leave knowing how to squeeze maximum throughput from your hardware for every backtest you run.
Prerequisites
Before you start tweaking agent settings, make sure you have:
- MetaTrader 5 build 2000 or newer — older builds have a less efficient agent scheduler. Check Help > About in the terminal.
- A multicore CPU — ideally 4+ physical cores. Hyper-threading helps but physical cores matter more.
- At least 8GB RAM (16GB recommended for large tick-data backtests). Each agent consumes memory; running 8 agents on 8GB will cause swapping.
- An Expert Advisor or indicator to test — you need something to run in the tester to verify your settings work.
- Administrator rights on Windows — agent installation may require elevation.
One note on MT4 vs MT5: MT4's Strategy Tester doesn't have local agents at all. It runs single-threaded by design. If you're on MT4 and want multi-core testing, you need to move to MT5. This guide is MT5-only.
Step-by-Step: Configuring MT5 Local Agents for Maximum Speed
Step 1: Open the Agent Manager
In MT5, go to View > Strategy Tester (or press Ctrl+R) to open the tester panel. At the bottom of the tester window, you'll see a tab labeled Agents. Click it. This opens the agent manager where you configure local and remote agents.
You should see at least one agent already listed — usually named "localhost" with a green status indicator. That's your default local agent, which runs on a single core by default. We're going to change that.
Step 2: Add Multiple Local Agents
Click the Add button in the agent manager. A dialog titled "Add Agent" appears. Here's what to set:
- Type: Choose "Local" (not Remote).
- Name: Something descriptive like "Core 1" or "Agent 2". This is just a label.
- Port: Leave the default (6433) for the first agent. For each additional agent, increment the port by 1: 6434, 6435, etc. This is critical — agents can't share the same port.
- Enable: Ensure the checkbox is checked.
Repeat this for each additional agent you want. A good rule of thumb: start with the number of physical cores your CPU has (not logical threads). For a 4-core/8-thread CPU, start with 4 agents. You can experiment with more later.
Step 3: Configure Agent Core Affinity (Advanced)
By default, Windows schedules agent processes across all cores. This usually works fine, but if you want finer control, you can pin each agent to a specific core using Task Manager. Here's how:
- Open the Agents tab in the tester.
- Start the tester once to spawn the agent processes (they appear as
metaeditor64.exeortester_agent.exein Task Manager). - In Task Manager, go to the Details tab, right-click each agent process, and select Set affinity.
- Assign each agent to a different physical core (e.g., Agent 1 to CPU 0, Agent 2 to CPU 1, etc.). Avoid assigning to hyper-threaded siblings (usually even-numbered CPUs on Intel).
This step is optional but can reduce cache contention on older CPUs. Most modern processors handle scheduling well enough that you can skip it.
Step 4: Run a Backtest to Verify Multi-Agent Operation
Open the Strategy Tester, select your EA, choose a symbol and timeframe, and set the testing period. Under the Settings tab, you'll see a dropdown labeled Agent. By default it says "All agents". Leave that — it tells the tester to distribute work across all enabled agents.
Click Start. Watch the Agents tab during the test. You should see multiple agents showing activity (they'll alternate between "Processing" and "Idle" status). The tester's progress bar moves faster than usual.
To quantify the improvement, run the same backtest with a single agent first, note the time, then run with multiple agents. The speedup won't be perfectly linear (there's overhead), but you should see a 3-4x improvement on a 4-core system.
Step 5: Adjust the Number of Agents Based on Test Type
Not all backtests benefit equally from more agents. The tester distributes work by splitting the date range into chunks. If your test period is short (e.g., 1 month of M1 data), the chunks are too small to benefit from many agents — the overhead of splitting and merging data outweighs the gain. For longer periods (6+ months of tick data), more agents shine.
Here's a practical guideline:
| Test Duration | Timeframe | Recommended Agents | Reasoning |
|---|---|---|---|
| 1-3 months | M1-M15 | 2-4 | Short period; overhead of splitting data limits benefit from more agents. |
| 6-12 months | M5-H1 | 4-8 | Good balance of data size and agent overhead. |
| 2+ years | Any | All physical cores | Large data sets benefit fully from parallelization. |
Tips and Best Practices from Experience
Over the years, I've tuned agents on everything from a laptop with a dual-core i5 to a 16-core Threadripper workstation. Here's what I've learned that doesn't make it into the official docs.
Don't Confuse Logical Cores with Physical Cores
Hyper-threading gives you two logical threads per physical core. Running 8 agents on a 4-core/8-thread CPU might seem smart, but you'll often see worse performance than with 4 agents. The reason: two agents competing for the same physical core's execution units cause cache thrashing. Stick to the number of physical cores, then test adding one or two more agents — sometimes the extra threads help with I/O-bound tests.
Use Tick Data Mode for Realistic Tests
If you're optimizing for speed, it's tempting to use "Every tick" mode for accuracy but "1 minute OHLC" for speed. The agent scheduler handles tick data much better than OHLC mode because the work units (chunks of tick data) are more uniform in size. OHLC mode creates uneven chunks that leave some agents idle while others grind through dense periods. For realistic backtesting with good parallelization, always prefer "Every tick" if your EA can handle it.
Monitor RAM Usage per Agent
Each agent loads its own copy of the symbol's tick data and the EA's internal state. For a complex EA with large arrays or file operations, a single agent might use 200-500MB. Multiply that by 8 agents and you're at 4GB just for the tester. If your system has 8GB total, Windows will start paging to disk, and your "fast" multi-agent test becomes slower than a single-agent run. Open Task Manager's Performance tab during a test — if memory usage is above 85%, reduce the agent count.
Disable Unnecessary Visualizations
When you're running multi-agent tests, the visual mode (the chart that shows trades being placed) forces the tester to synchronize all agents to a single visual display. This kills parallelism. Always run in non-visual mode for speed tests. You can enable visual mode for individual verification runs later.
Common Mistakes and Troubleshooting
Even with the right settings, things can go wrong. Here are the most frequent issues I've seen and how to fix them.
Mistake: Agents Show "Disabled" or "Error" Status
If an agent shows as disabled in the Agents tab, it usually means the port is already in use or the agent process crashed. Check that you're using unique ports (6433, 6434, 6435, etc.) and that no other application is using those ports. Run netstat -ano | findstr 6433 in Command Prompt to check. If a port is occupied, increment all your agent ports by 10 to avoid conflicts.
Mistake: Backtest Runs on a Single Agent Despite Multiple Agents Configured
This happens when the testing period is too short or the EA uses global variables that prevent parallel execution. Some EAs have logic that depends on sequential order — for example, reading/writing to a single file. The tester detects this and falls back to single-agent mode. Check the Journal tab in the tester for a message like "Cannot split the testing range into independent parts — using single agent." If you see that, your EA isn't safe for parallel testing. You'll need to refactor it to avoid shared state, or accept single-core speeds.
Mistake: Speed Improvement Is Minimal (Less Than 2x on 8 Cores)
Several culprits: your test period is too short (try a longer date range), your EA is I/O-bound (writing to files or databases during the test), or your CPU is thermal-throttling. For the I/O issue, minimize file operations inside the tester — batch writes and flush at the end. For thermal throttling, monitor CPU clock speeds with HWMonitor; if they drop below base clock during the test, improve cooling or lower agent count.
Mistake: Tester Crashes or MT5 Freezes with Many Agents
This is almost always a memory issue. On a 16GB system, don't run more than 6-8 agents for a tick-data test. If you have 8GB, stick to 2-4 agents. Also check that your antivirus isn't scanning the agent processes — add metaeditor64.exe and tester_agent.exe to the exclusion list.
Summary / Recap and Next Steps
Configuring MT5 local agents is the single most effective way to speed up backtesting, provided you understand the trade-offs. The key points:
- Add one local agent per physical CPU core, each on a unique port.
- Use "All agents" in the tester settings to distribute work.
- Match agent count to test duration — short tests don't benefit from many agents.
- Monitor RAM usage and CPU thermals during multi-agent runs.
- If your EA uses shared state (files, global variables), it won't parallelize — refactor or accept single-core speed.
Once you've confirmed your agent setup works, consider these next steps:
- Test with optimization mode — the agent scheduler handles optimization passes even better than single backtests, because each pass is independent. You can often run 16+ simultaneous optimization threads on an 8-core machine.
- If you need more power, set up remote agents on other machines on your network. The agent manager supports adding remote agents by IP address — you can distribute testing across a whole lab of computers.
- For serious development, consider a dedicated backtesting server with a high-core-count CPU (e.g., AMD Threadripper or Intel Xeon) and plenty of RAM. The agent settings scale linearly with cores.
Your backtests will never be the bottleneck again.
Frequently Asked Questions
How many local agents should I configure for my CPU?
Start with the number of physical cores (not logical threads). For a 4-core CPU, try 4 agents. For an 8-core, try 8. Then test adding one or two more — sometimes logical threads help with I/O-bound tests, but often they hurt due to cache contention. Monitor CPU usage in Task Manager: if any core stays below 90% utilization during a test, you can add more agents.
Can I run MT5 backtests on multiple computers using agents?
Yes. In the agent manager,
