What's new in ZenSearch

Agent runs now go through a five-layer budget system: complexity-aware tiers, per-sub-task wall-clock timeouts, a pre-flight dollar ceiling that rejects expensive plans before any tokens are spent, mid-run cost enforcement, and a soft-limit pause that saves a resumable checkpoint with a 7-day TTL when any budget exhausts.

After successful multi-tool sessions (≥5 tool calls, ≥0.7 synthesis confidence), agents now distill the workflow into a reusable procedure and progressively disclose it to future runs via the new view_procedure tool.

LLM-based observer extracts key findings, tool usage, and pending work from long sessions into a compact summary. Typically 80%+ token compression on 50+ message conversations; injected as a cacheable system-prompt prefix.

Timeouts, rate limits, and upstream 5xx errors are now auto-retried once before escalation. Retries don't count against tool-call budgets. Opt out via AGENT_RECOVERY_ENABLED=false.

cost_update SSE event streams every few iterations so users see real dollars ticking up against the per-run cap while the agent is still running. Cadence configurable via AGENT_COST_UPDATE_INTERVAL.

Gmail, Calendar, Drive, Docs, and Sheets — read and write. OAuth 2.0 or Google Service Account with domain-wide delegation. Agents now act as the end user, with per-user OAuth tokens.

OneDrive, Outlook, Teams, SharePoint, Planner, plus Graph-powered org-chart and calendar free-busy. OAuth through Azure AD with per-scope configuration.

Zendesk (14 tools — ticket CRUD, help-center articles, users, organisations), Airtable (8 — bases, records, comments), Notion (12 — pages, databases, blocks, comments). All via per-user OAuth where supported.

Admins can register internal HTTP endpoints as agent tools (name, URL, method, auth, parameter schema — encrypted credentials). For third-party tool servers, ZenSearch speaks Model Context Protocol (MCP) as a client — register an MCP server and its tools appear in the team's tool catalog.

Native assistants in both chat apps: Block Kit answers and approval cards in Slack, Adaptive Cards and threaded replies in Teams. Cross-surface continuity — the same conversation follows the user between web, Slack, and Teams.

Every AI call routes through a central gateway. Swap providers with one env var — Groq for speed, OpenRouter for unified billing across 100+ upstream models, Azure for per-tenant OpenAI, Bedrock for AWS-native Claude and Nova. Ollama still works via OpenAI-compatible endpoint for fully local deployment.

Anthropic cache_control ephemeral markers get a 90% discount on cached reads; OpenAI and Groq prefix caching get ~50%. Tracked as cached_input_tokens and cache_creation_input_tokens on the usage ledger.

ContentParts on chat messages lets agents pass image inputs (file uploads, extracted document images) to vision-capable models. Token estimation accounts for ~765 tokens per image part.

Content-hash dedup short-circuits unchanged documents before they enter the parse pipeline. A PageChangeDetector handles web-crawl delta. A DeletionDetector uses set-diff against each connector's seen source IDs to find removals, with a 50% safety threshold that refuses to act on suspiciously large deletion sets.

AI assistant in a browser side panel, aware of the current page context. Keyboard shortcuts, session persistence, and identity federation with the web app.

Agents produce structured outputs (code, reports, plans) into a side panel with diff view, version history, and user editing. Content types: markdown, code (Python, Go, TypeScript, etc.), JSON.

Composite confidence score (0.0–1.0) from coverage ratio, shape confidence, and violation-free signals; surfaced as High/Medium/Low confidence in the chat UI. Post-synthesis verification semantically checks citations against sources and corrects misattributions.

Natural-language-to-SQL over all four databases. Schema discovery, read-only query execution, and automatic error correction via LLM-friendly error messages.

Permission metadata imported from Confluence, SharePoint, Google Drive, and identity providers on every sync. Applied as a query-time filter — restricted documents never enter the ranking pipeline.

Prompt injection and PII detection on inputs; hallucination (lexical/semantic/hybrid), toxicity, and relevance checks on outputs. Configurable per team with hot-reload.