Agent Atlas / Coding agents

alejandro-ao/tau

A Python coding-agent harness whose readable architecture, extension seams, protocol changes, and contributor onboarding make the runtime itself a teaching object.

Coding agent Full dossier Reported + sourced Reviewed July 21, 2026
Intended operator

Developers learning from or extending a compact coding-agent harness.

Primary surfaces

Python CLI / agent harness / extension and protocol seams

Tracked repository

alejandro-ao/tau

Teaching mechanism

A protocol rewrite preserves yesterday's sessions

Compatibility code and tests keep older serialized message history usable while the wire and extension model changes.

Evidence limitFixture compatibility covers represented histories, not every external client or future migration.
Inspect the source

System Map

Eight inspection boundaries
Model boundary

Provider calls are wrapped by a readable Python harness rather than hidden behind a large application shell.

Context and memory

Messages, branch context, and persisted sessions expose what the loop carries forward.

Tools and execution

Python extension seams make tool and capability composition visible to contributors.

Permissions and isolation

The teaching surface identifies boundaries but does not by itself prove strong operating-system isolation.

Sessions and recovery

Protocol compatibility and serialized-history tests protect resumable work across rewrites.

Delegation

The inspected evidence is stronger on extensibility and sessions than on worker hierarchies.

Human control

CLI behavior and architecture documentation keep control flow legible.

Evidence surface

Small modules, examples, compatibility fixtures, and onboarding documents provide reusable inspection points.

Learn Through This Project

8 lessons

Compare Its Boundaries

6 labs

Latest Reporting

5 articles

Primary Sources

Open Questions

  • How far can a teaching-oriented harness grow without hiding the boundaries it was built to explain?

Evidence state describes what this Atlas profile can support today. It is not a product score and it is not evidence that uninspected capabilities are absent.