Agent
An agent that belongs to the mission.
The Excapia agent receives the mission, the relevant platform context, its tools, the evidence so far and its boundaries. Excapia, not the model, owns continuity, verification and termination.
The native execution loop
The agent works in steps that Excapia records, so progress survives restarts and every action stays inside the mission contract.
- ReceiveMission, contract, context and evidence so far
- PlanSteps toward the success criteria, within budget
- ActTools allowed by policy; each call checked
- CheckpointDurable progress; resume without repeating work
- VerifyIndependent assertions decide the outcome
Model inference and coding agents are different things
Calling a model API gives you text and tool calls from that model. It does not include a full coding agent such as Claude Code or Codex. Excapia treats them as two distinct capabilities.
Model-provider adapters
Excapia’s own agent loop calls a model you configure for reasoning, planning and tool selection. External fallback is an explicit deployment choice, never automatic.
External coding-agent adapters
For repository work, Excapia can delegate a bounded task to an external coding agent and treat its result as a claim to verify, like any other strategy.
Providers
Customer-managed endpoint, also used by AI Assist.
Self-hosted inference servers you run.
Anthropic API, if you choose to configure it.
Google API, if you choose to configure it.
OpenAI API, if you choose to configure it.
Not every model supports images, audio, video or tool calling. Availability depends on the model and configuration you choose.
Extending what the agent can do
Skills and experience
A Skills Library and experience captured from verified recoveries, retrieved by similarity when a familiar failure returns.
ExampleA selector repair approved last week is offered first when the same component changes again.
Plugins and Service Packs
Installable tools and domain packs, each declaring the actions and resources it needs so policy can govern it.
ExampleA payments pack adding read-only settlement checks.
Multimodal investigation
Screenshots, audio, and video analysed through sampled frames and transcripts, where the configured model supports them. Not full native video understanding.
Scheduling and bounded recovery
Built-in triggers and schedules, stall detection and a recovery pipeline that activates the agent only under configured conditions.
Approved disconnected work
Execution under an authority lease while a worker is disconnected, with journal reconciliation before new work.
Outcome verification
The agent’s “done” is a claim. Independent assertions decide whether the mission is verified.
Agent Studio
Watch the agent work, step by step
Conversation, task plan, repository, diff, terminal, tests and artifacts in one place, inside the mission’s boundary, ending in an approval request rather than a silent change.
