Dwi

Dwi by thienhocThe Human Layer for AI Agents

AI can move faster.You should
feel lighter.

Dwi gives agent work a clearer shape. It helps define the outcome, bound the scope, assign the right lane, control resource use and return evidence before you accept the result.

GitHub
See how Dwi works

It works around the agent tools you already use. No hidden runtime. No permission bypass. No need to install every module.

Open source · Six focused modules · Read, try and remove

Viewer lens

Choose a useful level of detail.

The system is doing more. Why does the work feel heavier?

Choose.Check. Decide.

Name a small job, see what was checked, then continue, correct, or stop.

Boundary

You know the work. The agent still makes it larger.

A small task can turn into new tools, broader plans and hours of review. Useful exploration quietly replaces the work you meant to finish.

Open the plain-language guide

Keep visible

Boundary
What this run may touch.
Proof
What was checked before moving on.
Your choice
Approve, correct, or exit.

Before / After: the same packet

The same case material is shown before and after it is arranged in a readable packet.

Before

  • Goal appears in one reply
  • Limit is buried in another
  • Proof arrives late
  • Decision waits for a manual search

After

  • Boundary is named once
  • Relevant facts travel together
  • Observed proof is marked
  • Human choice remains explicit

Human review stays separate from the packet.

The information does not become more certain by being rearranged; the packet simply makes its boundary, proof, and next decision easier to inspect.The same note is easier to read when the important parts are kept together.
Text equivalent
  1. Before: replies, limits, and pending questions are separated.
  2. After: one packet holds the boundary, observed proof, unresolved risk, and human choice.
Known workPartial subjectNew groundHidden work

A packet gives the work a shape

Give the work a shapebefore asking the agent to move.

A packet defines the outcome, context, allowed changes, off-limits areas, and evidence that must return.

Intent
What needs to become true?
Context
What does the agent need to know?
Authority
What may it read, change or decide?
Evidence
What must return before the result is trusted?

The work now carries more of its own memory.

Parallel work stays separate. The route back to one decision stays visible.

Context Packet Routing

Relevant inputs enter one packet on neutral routes, move to human review on a labeled review route, and use a solid route only after approval. Correction and manual exit remain visibly separate.

Route diagram from source material to a bounded packet and human review.

Human goalExisting factsWorking limitsContext packet{ }Human reviewApproved next step{ }Manual exitinputreview routeapprovedproposed / retryhuman-controlled exit
  • Human scope
  • Packet checkpoint
  • Observed route
  • Retry route
  • Input
  • Under review
  • Manual exit
Routes communicate state as well as direction: neutral is input, the review route is not yet approved, solid is observed or approved, dashed is proposed or a retry, and orange remains a human-controlled exit.Only the information needed for this job goes into one note before you look at it.
Text equivalent
  1. Human goal, existing facts, and working limits enter a single context packet.
  2. The packet moves to human review on a distinct review route.
  3. An approved route proceeds to the next bounded step.
  4. A proposed correction and a human-controlled manual exit remain visibly separate.

Use the strongest thinking where it matters.

A capable model can shape the plan, divide the work and judge the result. Lower-cost models can handle clear, bounded steps. Expensive reasoning stays focused on decisions, not repetition.

This is one valid Arc pattern. Arc does not require a particular model mix.

One plan. More than one useful lane.

Work that does not depend on the same files or decisions can move at the same time. Root brings compatible results together and keeps ownership clear.

Parallel work reduces elapsed time only when the lanes are genuinely independent.
See how the system worksRead the Arc caseSee how Arc works

One case: measured wall time

Confidence is not proof.

Dwi separates what was verified, observed, estimated, targeted or remains unknown. You do not have to decode how certain the agent really is.

Proportional evidence from one bounded case

Measured wall time is 1,930 seconds before and 1,176 seconds after in one case. The difference is 754 seconds, about 39 percent. Context reuse is 94.5 percent and effective-new context is 5.5 percent.

One case: measured wall time

Before1,930 seconds
After1,176 seconds

Difference in this case

754secondsabout 39%
These values describe one measured case. They do not establish a cause or promise the same result elsewhere.This is one timed example. It shows what was measured here, not what will happen every time.
Text equivalent
  1. One case: measured wall time was 1,930 seconds before and 1,176 seconds after.
  2. The difference in this case was 754 seconds, about 39 percent.
  3. Context reuse was 94.5 percent; effective-new context was 5.5 percent.
  4. The figures are descriptive, not a causal claim or universal benchmark.
CaseOne bounded caseComparedBefore and afterMeasured1,930s to 1,176sLimitNot universalEvidenceArtifacts and checks

One observed case. Not a universal performance promise.

Understand the evidence modelRead why polished answers are not proof

Six entry modules

Start with the pressure you feel.

Each module helps you make one part of the work easier to see before you choose the next step.

What this module changes

Conduct

Turn a conversation into an ordered next step.

Use one focused entry point first. The remaining modules stay optional.

Conduct

Starting out: Conduct or Lean. Already operating: Lean, Budget or Evidence. Operating professionally: Evidence first, then Bridge or Arc when justified.

Explore all modulesTry one module safely

The decision stays human

The decision stays human.

A safe system does not only know how to continue. It knows when continuing would exceed its authority.

Stopping at the boundary is correct behavior.

Safety / Manual Exit

When material risk is unresolved, escalation pauses and the case returns to human review. The platform's documented stop control remains available for a manual exit.

Safety route diagram ending in a human-controlled manual exit.

Material risk unresolved{ }Pause escalation{ }Return tohuman review{ }Manual exitKeep the platform's documented stopcontrol available
  • Human review
  • Packet checkpoint
  • Observed route
  • Manual exit
Manual exit is a visible route, not a hidden failure state. It remains available beside the staged workflow.If something important is still unclear, pause. Look at it with a person before you continue or stop.
Text equivalent
  1. An unresolved material risk pauses escalation.
  2. The case returns to human review.
  3. The platform's documented stop control remains available.
  4. A person can take the manual exit.
Read the operating rules

  • Pause escalation when a material risk is unresolved.

  • Return the case to human review before reopening the route.

  • Keep the platform's documented stop control available.

The system can recommend. You decide what becomes real.

Read the safety modelExplore the project

Behind Dwi

Built from practice. Shared for others to test.

Dwi is an independent project by Trần Thiện Học. It grew from more than a decade of work across strategy, communication, digital experience and complex systems, followed by sustained hands-on use of AI agents. The project is shared as an open research preview. Each module is intended to remain small enough to read, try and remove.

Read the storyAbout the researchSee the evidence policy

Begin with one pressure point.

Choose one module. Try it on a reversible task. Keep it only if the work becomes clearer.

Choose a moduleRead the safe trial guide