T-Chat Harness 2.0
More autonomy.
Clear controls.
The model proposes and reasons; T-Chat governs tools, permissions and checks. Choose how work progresses without turning autonomy into unlimited authorisation.
Benchmarks are comparison tools, not a general certification of a model or the app. Keep human review for important outcomes.
Agent management
Choose when to step in
Set autonomy at project level. The agent can ask for missing information and continue after your answer; sensitive actions still follow permissions and confirmations.
- Planning, model reasoning and review are separate controls.
- Confirmation requests describe the action to be authorised.
- Generated text or an external document does not replace your consent.
Continuity and traceability
Return without losing the process
Task steps, status and results remain recoverable when you change sections. Processing time and technical data help distinguish ongoing activity from a completed result.
- Copy the full agent process for support and review.
- Checks use actual tool results, not just the model’s claims.
- Errors and blocked actions stay visible; a passed check does not guarantee infallibility.
Benchmark lab
Measure on your hardware
Compare models and configurations with repeatable tests, exportable results and synthetic data. The lab combines application checks with local inference measurements.
- Loading, first-token and generation times, plus cache data when exposed by the runtime.
- Comparisons between runs and checks on results, not speed alone.
- One benchmark at a time; tests use resources and may slow other work.
Inference profiles
Balance speed, memory and quality
Adjust context, reasoning and model residency in memory. Profiles align options across supported workflows; benchmarks help assess the trade-off on real work.
- Adaptive context within configured limits and cache statistics kept distinct from estimates.
- Configurable summarisation model, with checks on retained information.
- Options such as Q8 cache depend on the backend: less memory does not automatically mean more speed.
Frequently asked questions
Is a higher reasoning level always better?
No. It can increase time and resource use without improving a simple request. The chat selector takes precedence over the profile; Model profile uses the configured setting. Compare correctness and duration before changing defaults.
Does autonomy mean no confirmations?
No. Authorisation remains in force. Asking the agent to proceed does not automatically authorise sending, publishing, deleting or accessing resources beyond the task.
Does caching speed up every model?
Context reuse can avoid repeated work, but it depends on the engine, model and prefix stability. T-Chat distinguishes measured data from unavailable information and does not promise a universal performance gain.
Use cases
Start with a real process.
Choose an activity, prepare a non-sensitive data sample and let’s define the expected outcome, hardware and checks together.