A new model ships almost every month. Each time, we hear from teams who wonder if they should rebuild their AI plans around it. Usually, they don't need to. The model is the part of an AI strategy that changes fastest, which makes it the worst part to build a plan around.
Here are four lessons from teams that build and run AI agents with d5s, and from running our own. Each lesson is also part of an AI strategy: a company's plan for which work AI does, under which rules and at what cost.
- Keep approvals in the runtime. A prompt can ask a model to wait for approval. Only the runtime can make it wait. In our test of five open-source LLMs, two would have sent an email without it.
- Build agents the whole team shares. A shared agent does a repeated job once for everyone, and each run starts with more context than the last.
- Make the model a setting. Teams that keep the knowledge with the agent change models without starting again.
- Track cost per agent. A model bill shows what you spent. Cost per agent shows what the money paid for.
1. Keep approvals in the runtime, not the prompt
Put every send and change behind an approval that the runtime enforces. In our test of five models, two would have sent an email without approval if only the prompt had stopped them.
In our conversations, trust came up before any question about models. And for good reason: a prompt can ask a model to wait for approval, but it can't make it wait.
We tested this ourselves. We gave five open-source models the same email task, with one rule in the prompt: ask for approval before you send. We left out the runtime gate on purpose, and two of the five tried to send anyway. The full results are in Best open-source LLMs for AI agents in 2026.
In d5s, actions that send or change data are set to Ask first by default. The runtime holds the action and shows an approval card. Nothing goes out until a person approves it, whatever model runs the agent.
- Approve & sendThe email goes out as written.
- RejectNothing goes out.
The person sees the recipient and the content before anything goes out, and can reject the send if something is wrong. The rule holds for every model, so switching models doesn't reopen the question of trust. See how d5s keeps a human in the loop for every send.
2. Build shared agents, not personal chat tools
Give a repeated team job to one agent that the whole team shares. The team pays once to build its context, and every run after that starts with more of it.
The teams we spoke with got the most out of agents that own a team job: sales follow-up, the finance close, support triage. People work with these agents in d5s, Slack, Microsoft Teams and email.
- d5s
- Slack
- Microsoft Teams
- Instructions
- Memory
- Playbooks
- Connected tools
Shared agents also save tokens. When five people each ask a personal chat tool for the same weekly report, the model rebuilds the same context five times: the same files, the same lookups, the same instructions. A shared agent does the work once and saves the result where the whole team can reuse it.
The benefits compound. A shared agent keeps useful context in memory and updates its playbooks from feedback. Each run starts with more context and needs less coordination than the last. A personal chat starts from zero every time.
That shared context belongs to the team, not to the model. It is the core of an AI strategy, and it must survive a model change.
3. Treat the model as a setting, not the strategy
Treat the model as a setting on each agent. When a better model ships, change the setting and keep everything else.
Teams told us they changed models more often than they planned to. In d5s, the model is a setting on each agent. The instructions, memory, playbooks, tools and approval rules stay when it changes.
A model change shouldn't be guesswork. Every night, our agents do real tasks with real tools, and independent judges check each outcome. We run new models through the same tasks, so we know the runtime still works with them before you switch. Read how our agent swarms help test d5s.
Most teams also found that no single model was best at every job. A fast model suits quick, frequent work. A more accurate model suits careful work where an error costs more.
4. Measure cost per agent, not per model bill
Measure cost for each agent, not for each model. Then you can tune the agents that use the most.
A model bill tells you what you spent. It does not tell you which work the money paid for. In d5s, each run is charged to the agent that did it. Agent stats shows runs, tokens and spend per agent, and a captured trace shows each model call and action in a run.
| Agent | Runs | Spend |
|---|---|---|
| Support triage | 1,240 | $186 |
| Sales follow-up | 410 | $74 |
| Finance close | 36 | $52 |
With cost per agent, a team can make one decision at a time: move the busiest agent to a cheaper model, or keep the accurate one where errors cost more.
How to build an AI strategy that outlasts the model
These lessons point to a short checklist.
- Put the rules in the runtime. Set each action to Allow, Ask first or Block. Rules in a prompt change with the model.
- Start with a shared job. Pick work that more than one person repeats, and give it one agent.
- Keep the knowledge with the agent. Instructions, memory and playbooks belong to the team, not to the model provider.
- Choose the model per agent. Match speed and accuracy to the job, and change the model when your own tests say so.
- Measure cost per agent. Find the agents that use the most, then tune those first.
What this means for your team
An AI strategy is the rules, the shared agents and the cost tracking around the model. In d5s, each agent keeps its knowledge and approvals while you change the model it runs on. A new model becomes a settings change, not a new project. Read how d5s runs any model on any cloud.
Questions and answers
Should an AI strategy depend on one model?
No. A new model ships almost every month, and no single model is best at every job. In d5s, each agent keeps its instructions, memory, playbooks and approval rules when its model changes, and d5s tests new models on the same real tasks as its nightly evals.
How do you stop AI agents from sending the wrong thing?
Put approvals in the runtime, not the prompt. In our test, two of five models would have sent an email without approval if only the prompt had stopped them. In d5s, sends and changes are set to Ask first by default, so a person approves or rejects each one.
Why are shared AI agents cheaper than personal chat tools?
They do a repeated job once for the whole team. Personal chats rebuild the same context for each person. In d5s, a shared agent keeps memory and playbooks, so each run starts with more context than the last.
How do you measure the cost of AI agents?
Per agent, not per model bill. d5s charges each run to the agent that did it. Agent stats shows runs, tokens and spend for each agent, so a team can tune the busiest agents first.
