Agent capabilities
Axis exposes harness capabilities to agents. Exact tool names and schemas can evolve; the stable customer model is the job each capability performs and the evidence it must return.
Real Axis UI with deliberately seeded, privacy-safe demonstration content.
Context and learning
Section titled “Context and learning”Agents can recall relevant memory before rediscovering decisions and save a deliberate lesson when it occurs. They should use AGENTS.md for standing rules, project/source state for current truth, and Memory for learned knowledge.
A memory result should include provenance and confidence where useful; recalled mutable facts still need live verification.
Planning and unfinished work
Section titled “Planning and unfinished work”Agents can discover or continue a shared work plan, add verifiable items, mark the active leaf, record evidence, defer future work, and find due or forgotten items.
They should not create a duplicate plan merely because a new session began. Clearing one local view must not silently destroy other sessions’ shared items.
Session continuity
Section titled “Session continuity”Capabilities can title the current session, set its status, resume an exact existing session, start a fresh independent session, or create focused child sessions.
These operations are different. Resume preserves exact durable identity. A new session creates a separate thread. Sub-agents create children under a coordinating parent.
Human judgment
Section titled “Human judgment”Agents can present structured questions and request approval for consequential actions. A gate should appear in the owning session and Inbox, state the exact blocker and scope, and stop the action until the answer arrives.
A notification that input is needed is not itself the decision surface.
Workspace and services
Section titled “Workspace and services”Agents can inspect the current workspace, reuse or open session-owned panes, register a durable saved result, and launch a named long-running service.
Inspection comes first so the agent does not duplicate an authenticated browser, live preview, or canonical service. Background work should attach quietly rather than steal focus.
Browser and visual proof
Section titled “Browser and visual proof”Local browser capabilities can navigate and interact with a real page, read state, and capture screenshots. Sensitive typed values and credentials must not appear in the human trace. Browser-visible work is complete only after the real state is opened, captured, inspected, and compared with the claim.
Routines and continuation
Section titled “Routines and continuation”Agents can create/manage Routines or quick actions where supported and schedule bounded check-ins for exact sessions. Long-running continuation must stop on completion, cancellation, handoff, or a genuine human blocker; it is not an invisible infinite loop.
Results are part of the contract
Section titled “Results are part of the contract”Every capability needs a readable success, empty, loading, and error presentation. Summary comes first; exact evidence is disclosed progressively; raw output is a final fallback. Secrets and high-risk identifiers must be redacted.
A tool transport acknowledgement is not proof that the intended side effect occurred. Verify writes through authoritative persisted or destination state.