Hexabot

Hexabot's real differentiator is that workflow logic is code you can version, diff and review — not a black-box canvas. Definitions live as YAML on workflow versions with three required sections (defs, flow, outputs) and an optional defaults block whose settings are deep-merged into every task; task-level settings override defaults and undefined values don't erase them. That gives you a clean inheritance model for timeouts and retries across a large graph. Execution settings are genuinely operational: timeout_ms of 0 disables the timeout wrapper entirely, and retries support enabled, max_attempts (default 3), backoff_ms (default 25), max_delay_ms (default 10000), jitter and multiplier — the kind of knobs you want before pointing an automation at a flaky third-party API. Expression handling is deliberately shallow rather than recursive: for task inputs only each top-level value is compiled, so a dynamic object has to be expressed as one top-level expression that returns the object. That is a design constraint worth knowing before you start, because it changes how you structure inputs. AI capability comes through binding kinds rather than magic: model bindings mount a language model provider and model ID and forbid actions, while memory bindings can be mounted more than once and expose mounted memory values into $context. Retrieval is pluggable — the built-in database full-text helper, a vector helper for SQLite or PostgreSQL, or your own custom helper against any retrieval system — which is the pragmatic choice for teams whose content already lives somewhere specific. The costs are real and should be priced in. You run Docker, you own upgrades, secrets, and channel credentials, and any CRM, ERP or marketing-suite connector you need is something you build as a custom action or channel. There is no managed cloud to fall back on, and no one to call when a channel token expires at 2am. Best fit is a developer-heavy team or an operator with engineering support, especially where compliance or data-sovereignty rules rule out sending conversation data to a hosted platform. If your team has no one to own infrastructure, this will stall.