Skip to main content
TRW
Skip to content
TRWWhy an Engineering Operating Layer for AI Agents

Why TRW

TRW is useful when the expensive part of AI-assisted engineering is no longer generating code, but preserving decisions, resuming work, and proving what actually passed. It makes those records explicit; it does not make every session remember everything automatically.

Start with three evaluation questions

  1. 01Which project facts are repeatedly rediscovered or contradicted?
  2. 02Which validation and review evidence should survive the conversation that produced it?
  3. 03Can a fresh session resume from repository records without trusting stale chat context?

The observable problem

A client or model may already import repository instructions and read source effectively. The gap appears when useful state exists only in a closed conversation or an unsupported completion claim. Look for these failures in your own work rather than assuming a fixed savings estimate.

Rediscovery

A later agent spends time recovering a decision or gotcha that was known but never stored in a searchable project record.

Unverifiable completion

A fluent summary says work passed, but the command, scope, counts, or review receipt are missing.

Interrupted execution

Context compacts or a session ends, leaving no checkpoint that another agent can inspect and resume.

Ambiguous handoff

Multiple contributors return prose instead of explicit ownership, changed paths, dependencies, and integration evidence.

Claim, mechanism, evidence, boundary

Evaluate TRW by the records it can produce, not by a promise that agents become smarter. Each need below has an inspectable mechanism and a limit.

Reuse a project findingInspect
Mechanism
Record it with trw_learn; retrieve it through session start or focused recall.
Evidence
A stored entry with detail, tags, source context, and an explicit recall result.
Boundary
Recall returns a candidate to verify. It does not prove universal task improvement.
Resume interrupted workInspect
Mechanism
Create a run and checkpoint meaningful milestones.
Evidence
Active-run state and the latest checkpoint under .trw/runs/.
Boundary
The checkpoint can be stale; compare it with current source before continuing.
Make completion inspectableInspect
Mechanism
Run project-native checks, then record their real outcome with trw_build_check.
Evidence
Command output plus scope, counts, failures, coverage when measured, and static-check state.
Boundary
The receipt is caller-reported. TRW records checks; it does not execute them.
Apply a hard delivery policyInspect
Mechanism
Evaluate task type and recorded evidence at trw_deliver.
Evidence
A passing receipt, structured acceptable-failure record, or authorized override.
Boundary
Phase guidance and optional hooks are not substitutes for the server-side gate.

Evidence order

When speed and certainty compete, use this order to decide what must survive the handoff.

  1. 1. Current truth

    Read the live repository and report skipped, failed, partial, or unverified work explicitly.

  2. 2. Project-native proof

    Prefer the repository’s commands, artifacts, and acceptance criteria over a plausible narrative.

  3. 3. Reusable records

    Preserve only findings and handoffs that a later agent can inspect, verify, update, or retire.

  4. 4. Proportionate process

    Use the smallest ceremony tier that still includes the evidence the task’s risk requires.

Product boundaries

Not a model or IDE

TRW works above supported coding clients. It does not generate code or replace the editor and model you choose.

Not automatic conversation memory

Agents explicitly record and recall selected state. Chat content is not silently promoted into future instructions.

Not a test runner

Your project runs its own tests, build, lint, and type checks. TRW records the supplied result and evaluates policy.

Not an unconditional zero-egress guarantee

Project memory is local by default, and memory_local_only=true denies memory network access. Auth, downloads, hosted features, or explicitly invoked network tools have separate boundaries.

Run a project evaluation

Choose one active repository and compare similar work before and after adoption. Define the observation window by enough repeated tasks to expose the problem; do not assume a universal setup time or payback date.

RediscoveryInspect
Baseline
Count repeated explanations or re-solved gotchas in comparable tasks.
Useful signal
A relevant stored record is recalled and verified before the work repeats.
RecoveryInspect
Baseline
Interrupt a real multi-step task at a natural milestone.
Useful signal
A new session can explain and resume the checkpoint from current source.
Claim accuracyInspect
Baseline
Compare completion summaries with CI and review evidence.
Useful signal
Recorded scope and outcomes match the authoritative project checks.
Operating costInspect
Baseline
Measure the calls, review effort, and maintenance your team actually adds.
Useful signal
Recovered context or avoided rework is worth the observed process overhead.

When to skip it

Throwaway work

If no later session will reuse the code or evidence, durable project records add little value.

No recurring context

A small uniform project may not have enough non-obvious decisions or interruptions to justify another layer.

No willingness to record evidence

Optional hooks cannot create trustworthy receipts on their own. If nobody will record real outcomes, the framework becomes decorative.

Incompatible repository policy

If project-local .trw state or the required client integration is prohibited, resolve that operating constraint first.

Next

If the premise lands, the quickstart gets TRW installed and wired to your AI client. If you want the mental model first, start with core concepts.