01
Direct response
One useful answer with no external action.
- Draft an email
- Summarize supplied notes
- Explain a concept
Agent workbench
Plan the work. Then give AI room to act.
Good results now depend on more than a clever prompt. Tangerine helps you choose the right way to work, define what an agent can do, and understand the context and tokens involved.
- Useful with any model or platform
- Designed for agents, APIs, and real workflows
- Runs locally in your browser
Choose by the shape of the work
No model requiredDo not classify the wording. Ask what the system must do, how often it must run, and whether the work can be divided.
02
Single agent
One worker completing several steps with tools.
- Research current sources
- Update one product feature
- Analyze and verify a file
03
API workflow
Repeatable AI work inside a product or process.
- Classify support messages
- Extract invoice data
- Create a scheduled report
04
Multiple agents
Independent workstreams with one shared outcome.
- Build a new product
- Compare several markets
- Plan campaigns by audience
Durable guidance
A task becomes agentic when it can act.
Prompts still matter, but they are only one part of the system. Reliable agents need a clear job, well defined access, and proof that the work was completed.
Set a task boundary
Give the agent one outcome it can finish. State what is in scope, what is out of scope, and where it must stop for a decision.
Treat tools like contracts
Define what each API or tool accepts, what it returns, and what can fail. Use the smallest set of permissions the work requires.
Close the evidence loop
Do not stop at a confident answer. Require checks, source links, test results, or another form of evidence that matches the task.
Agent contract
Five parts of a task an agent can finish.
Use this structure when a job moves beyond conversation and into files, services, or decisions.
Outcome
Name the result, who needs it, and the form it should take.
Context
Supply the sources, definitions, examples, and durable instructions that shape the work.
Capabilities
List the files, APIs, search, code execution, and other actions the agent can use.
Boundaries
Define permissions, budget, time, approval points, and actions that are never allowed.
Evidence
Decide what proves the result is accurate, complete, and ready for a person to review.
API thinking
The model is one part of the product.
A dependable AI feature also needs state, tools, permissions, logs, and a useful way to recover when something goes wrong.
Tools are contracts
Use clear inputs, predictable outputs, useful error messages, and narrow permissions.
State has to live somewhere
Choose what belongs in the current request, the product database, or durable instructions.
Autonomy is a choice
Let low risk work proceed. Ask for approval before actions that affect people, money, or production systems.
More agents mean more coordination
Parallel work can help when tasks are independent. It also creates handoffs, conflicts, and more context to manage.
Logs are part of the experience
Record decisions, tool calls, failures, and final evidence so a person can understand what happened.
Test real workflows
Evaluate complete tasks with realistic tools and messy inputs. A polished answer alone is not enough.
Context lab
See what fills the working window.
Tokens carry instructions, source material, tool results, history, and the final response. Adjust the mix to see how context complexity changes.
4
3
3
6
1
Context complexity
62
Growing
An educational comparison, not a measurement of actual token usage or model limits.
This mix is workable. Keep the instructions stable and pass only the source material needed for the current task.
Token basics
Tokens are fuel, not the destination.
The goal is not to use the fewest tokens. It is to spend them on context and evidence that improve the result.
Context window
The working limit includes instructions, messages, source material, tool results, and the response being produced.
Input and output
Input tokens carry the task and its context. Output tokens carry reasoning artifacts, tool instructions, and the final result.
Useful efficiency
Remove repeated or irrelevant material. Keep the facts, constraints, and examples that help the system make better decisions.
One project, four routes
“Build a new product” is not one kind of task.
The route depends on what you actually expect AI to do. Start with the smallest route that can finish the real job.
Direct response
Use it to explore the idea, compare possible architectures, name the product, or draft a project brief.
Single agent
Use it for a focused feature when one agent can inspect the code, implement the change, and run the checks.
API workflow
Use it when AI is part of the product itself and must respond to users, data, or system events reliably.
Multiple agents
Use it when research, design, front end, back end, and quality checks can proceed as distinct workstreams.
Common questions
A practical starting point.
An agent can pursue an outcome across more than one step. It can inspect context, choose actions, use tools, observe the results, and continue until the work is done or approval is needed.
Use an API when the work needs to run inside a product, follow a repeatable process, connect to other systems, or operate at a scale that manual conversation cannot support.
They help when parts of the work are independent and can be reviewed separately. They add less value when every step depends on the previous one or when coordination costs more than the time saved.
Put stable rules there: how work should be handled, what must never happen, what quality means, and which preferences apply across many tasks. Keep temporary facts with the task itself.
No. More context can preserve useful evidence, but unrelated history and repeated instructions make the important material harder to find. Relevant context matters more than raw volume.
Privacy note: The context lab runs in this page. It does not call a model, upload data, or store your changes.