Skip to main content

Agents

Build AI agents with tool use, function calling, MCP servers, and streaming step output.

📄️ Agent Swarm

Market-based orchestration over an AgentFederation. Instead of round-robin or capacity routing, AgentSwarm runs an auction for every task: candidate agents submit bids describing estimated cost, quality, and confidence; a BidStrategy scores those bids against each agent's learned Reputation; and winners are selected by the configured AuctionType. Outcomes feed back into reputation (via UCB or Thompson sampling), so agents that consistently deliver win more of their category over time. Runs are deterministic when you set a seed.

📄️ Self-Improving Agent

Wraps any agent in an eval-gated self-improvement loop. SelfImprovingAgent observes production feedback, has a MetaAnalyzer propose signed, reversible AgentConfigPatch diffs (prompt rewrites, tool changes, routing rules, few-shot examples), validates each candidate against an EvalSuite (or PromptOptimizer), and only promotes it to a canary rollout through an AutoRolloutManager if it clears the eval gate. Every patch is signed, appended to an append-only audit log, and can be rolled back to its recorded before snapshot.

📄️ Computer Use Agent

Drives a screen with a provider-agnostic observe → plan → act loop. ComputerUseAgent captures screen state through a ScreenProvider, asks a ComputerUseProvider for the next step, normalizes every provider's output into a single ComputerAction schema, runs it through a conservative SafetyPolicy, and executes it — pausing for human confirmation on risky actions. The same agent code runs against Anthropic computer-use, OpenAI's computer tool, or an open-source local model, and can record every session to JSONL for exact replay.

📄️ Neuro-Symbolic Agent

Pairs an LLM with a formal solver. The LLM reads a natural-language problem and extracts formal constraints; a symbolic backend (Z3, SymPy, MiniZinc, or Prolog) solves and independently verifies them; and the LLM only renders a final answer once the solver has confirmed a satisfiable model. Every run carries a ProofTrace describing the solver status and model, so answers are checkable rather than merely plausible. Unverified answers are handled by an explicit on_unverified policy — retry, reject, or flag.

📄️ Digital Twin Agent

Learns your authentic writing voice and drafts commit messages, PR descriptions, and code reviews that sound like you — never auto-sending anything you have not gated as safe. DigitalTwinAgent distills your text samples into a versioned StyleProfile (tone, structure, vocabulary, review style), generates drafts through your LLM, and scores every candidate with a VoiceMatcher (n-gram overlap, vocabulary match, structure match, plus an optional LLM judge). An enforced DelegationPolicy gates dispatch: neversendauto channels are hard-blocked, and draftwithapproval channels refuse to send without an explicit human approval token.

📄️ Time-Travel Codebase

Reason across a repository's entire evolution — not just its current state. TimeTravelAgent wraps a git repo behind a GitBackend, parses commit history into an AST-based EvolutionIndex (symbol- and file-level change entries with #NNN PR linkage), flags stale abstractions with a DriftDetector, and turns diff timelines into readable markdown with a DiffNarrativeGenerator. You can scope any query to a point in time with agent.as_of(date), ask natural-language questions about how code got the way it is, and get an LLM (or heuristic) narrative back.

📄️ Signed Agent Marketplace

Package an agent as a portable, cryptographically signed .agent bundle, verify it end to end, install it safely, and share it through a self-hostable registry with signed reviews and eval-based ranking. A bundle is a deterministic ZIP whose manifest.json inventories every file with a per-file SHA-256 digest, and whose manifest is signed with an Ed25519 publisher key. Nothing in a bundle is ever imported or executed during verification or installation — an installed agent is inert and can only run once you attach an explicit sandbox.