An always-on operator for your paid media.

Yes. The safety model is the same one your in-house team operates under, made stricter by guardrails. The agent connects via Meta and Google's official APIs (the same OAuth flow you use for any reporting tool). You set spend caps, daily-change ceilings, allow-listed campaigns, and escalation thresholds. The agent cannot touch anything outside the rails you draw. Every action is logged with the trigger and is one-click reversible. SOC 2 Type II.

Three protections, in order. Guardrails: the agent can't take actions outside the rules you set (max bid changes per day, no single creative pause without 95%+ confidence, etc.). Logging: every action carries the trigger, the supporting signal, and a confidence score. Reversibility: any move can be rolled back in one click, individually or as a batch. Most customers configure auto-pause if any 24h window underperforms a baseline by >X%.

Brand voice rules, sensitive-category vetoes, and creative blacklists are configured by your brand lead and live in Feather as guardrails the agent must respect. Final brand-pass on launch and flagship creatives is gated by a human (see Human-in-the-Loop). The agent will not generate, push, or scale creative outside the brand corpus you've defined.

Your account data lives inside Feather, in your environment. .feather is one embedded file. No third party in the loop, no data leaving for training. We do not train foundation models on customer data. Aggregated, anonymised pattern signals (e.g. "off-peak bid lifts work in IN D2C") may flow to the broader pattern library only with explicit opt-in.

Always. Every action shows up in the Decision Log within seconds. You can pause the agent globally, pause it per campaign, or roll back any single move. Many customers run a "shadow mode" for the first 2 weeks where the agent suggests every move but a human confirms. Once trust is built, they switch to autonomous on a per-channel or per-campaign basis.

Yes. That's the sweet spot. The bigger the spend, the more leakage from manual ops, and the more the agent's 24/7 coverage compounds. Customers in the $1M to $5M/month band typically see the largest absolute ROAS lifts, because percentages compound on bigger denominators. Spend caps and per-channel caps make this safe.

No. It changes their job. The manual lifecycle (variant generation, A/B execution, fatigue detection, spend reallocation, hypothesis testing, competitor scraping, reporting) moves to the agent. Your creative directors, brand leads, and performance leads stay in the seats that matter most: narrative direction, brand voice rules, channel mix, escalation thresholds, quarterly strategy, final brand-pass on flagship creatives, and approval of new audiences, geos, or partnerships.

See the Human-in-the-Loop section above. As a rule of thumb: agents do the labor; humans do the judgement. Anything tactical, operational, repeatable, and bounded by rules → agent. Anything strategic, brand-facing, or involving new categories of risk → human.

You still need strategy, brand, and oversight. You need fewer hours of routine ops. Most customers retain a senior performance lead who acts as the agent's manager: setting strategy, reviewing weekly, approving new audiences. Junior buying roles often shift toward strategy and creative oversight rather than dashboards. Agencies that move toward strategic and creative work continue to add value. Pure execution agencies are the ones replaced.

You can, and a few teams do. The build cost is rarely the LLM API. It's the ad-platform integrations (Meta, Google, YouTube each have non-trivial quirks), the safety layer (guardrails, escalation, audit logs, rollback), the memory layer (patterns that decay, stickiness, propagation across geos: Feather), the evaluation harness, and the on-call. Most teams that try this discover they've taken on 6 to 9 months of engineering work that doesn't differentiate their product, just to reach feature parity. Hawky exists so you don't have to.

Yes, for enterprise customers. Feather is embedded by design (one file), so the substrate runs in your environment with zero outbound. The agent itself can run in your cloud or as a managed deployment. Talk to sales for the deployment options that fit your security review.

The short version: most of those tools generate creative variants or surface optimization recommendations. They don't do the work. They hand a recommendation back to a human. Hawky is end-to-end: it generates, tests, runs, optimizes, scales, and logs. We have a longer competitive teardown. Happy to share on a call.

No. Connection is OAuth-based, so your performance lead can do it in 15 minutes for Meta, Google, and YouTube. If you want custom guardrails or escalation hooks (Slack, email, custom webhook), you'll want an engineer for an afternoon. Fully Managed customers don't touch setup at all.

Managed customers typically see directional signal in week one (decisions logged, fatigue caught, overnight wins captured). DIY customers see meaningful KPI shifts inside 30 days. Most customers see ROAS uplift inside 90 days. There's no warm-up tax. The agent learns from your historical context the moment Feather is connected.

Cohort median across 200+ active customers: +25% ROAS in the first 90 days, −43% CAC on competitive categories, +10× hypothesis testing velocity. Range is wide. Accounts with poor manual hygiene see larger lifts. Well-optimized accounts see smaller percentage lifts but larger absolute gains. We measure against your own pre-Hawky baseline, not industry benchmarks.

Through Feather's Living Context. Every test outcome, every winning creative, every audience that worked or didn't gets captured, decayed by recency, and propagated to similar accounts or geos when a pattern is statistically meaningful. Week one the agent is good. Week twelve the agent has 600+ patterns specific to your brand and is dramatically better than week one. It compounds.

You keep your data, your account state, your decision log, and a Feather export. The agent disconnects from your ad accounts. No lock-in, no data hostage. We'd rather you stay because it's working than because leaving is hard.

A monthly subscription minimum gets the agent operating. The larger share is tied to your KPI moving — when your ROAS climbs, we earn more. We share specifics on a 30-minute call after we know your spend, channels, and KPI definition.

Yes. Most enterprise engagements start with a 30-day pilot on one account or one channel. We agree the success metric upfront, run the pilot, and the pilot fee converts into the main engagement if you continue. If we miss the metric, the pilot is on us.

DIY: your team operates the agent. Best for in-house performance teams that want speed without giving up the wheel. Fully Managed: our growth pod operates the agent on your behalf. Strategy, guardrails, weekly reviews, monthly outcome reports. Best for teams that want outcomes, not operations. You can switch between modes without re-onboarding.

15 minutes to connect the ad accounts. 24 hours for Feather to ingest 18 months of historical context. 7 days for the agent to ramp into autonomous mode (or longer in shadow mode if you prefer). Most accounts have meaningful agent-driven wins logged inside the first week.

Feather is the Living Context layer the agent stands on. CDPs are static (they snapshot the world). Warehouses don't decay (yesterday's truth weighs the same as today's). Vector databases don't know your brand. Feather is a living substrate. Patterns are tagged, indexed, decayed by recency, and propagated where they matter next. It's MCP-native, embedded as one file, and your data never leaves.

Yes. Feather is its own product. Hawky is its first consumer. Other agents (and your own internal agents) can plug into Feather as the substrate. If you want to talk to us about Feather independently, we're happy to.

Meta, Google, YouTube on day one. The rest of the perf-marketing stack (TikTok, LinkedIn, Reddit, Apple Search Ads, programmatic) is on the roadmap. Living Context preserves account memory across all of them as they come online, so there's no re-learning when a new channel is added.

The agent reads attribution from the platforms you already use (Meta CAPI, GA4, your MMM if you have one). It does not make attribution decisions for you. It operates against the KPI you report on, with the attribution model you trust. We can also plug into your own warehouse for blended ROAS calculations if you have one.

Feather is the Living Context layer that AI agents plug into for performance marketing. It's a living substrate, meaning the data has decay, stickiness, and multimodality built in. Shipped as one embedded file, MCP-native, where your data never leaves your environment. Hawky's Performance Agent reads from and writes to Feather. Other agents (yours included) can plug in the same way.

A traditional database stores rows. A vector DB stores embeddings. A warehouse stores history. None of them decay the way performance signal does. Yesterday's CPM win matters less than today's, and last quarter's pattern only matters if it keeps mattering. Feather treats memory as living: stickiness for things that keep working, decay for things that don't, multimodality (text, visual, audio) under one entity ID. It's not a better Postgres or pgvector. It's a different category.

None of those, and that's the point. Vector DBs are commodity infrastructure. Feather operates above them. CDPs are static. Feather is living. Warehouses don't decay. Feather does. If you slot Feather into one of those buckets, you'll mis-evaluate it. The closer mental model is "Performance Context Layer": purpose-built substrate for the way agents actually work.

Yes. Feather is MCP-native (Model Context Protocol, Anthropic's open standard). Any agent that speaks MCP can read and write Feather as a tool, not a dataset. That includes Claude, your own Claude Agent SDK builds, GPT via MCP shims, and internal tooling. Hawky is the first consumer. Other agents are the next.

It means Feather presents itself to your agents as a set of tools, like `query_audience`, `recall_winning_creative`, `get_brand_voice`. Each one has typed inputs and outputs, the same way the agent calls the Meta API or sends a Slack message. You don't write SQL. You don't manage embeddings. The agent treats Feather like any other tool in its toolbelt, and Feather handles decay, indexing, and propagation underneath.

Feather ships as a .feather file that lives inside your environment. No SaaS round-trip. No "send your performance data to a third party for inference." This matters for two reasons: sovereignty (your data never leaves, nothing trains on it, auditing is trivial) and speed (queries don't pay network latency, so the agent gets sub-millisecond context lookups).

Yes. Feather is its own product. Hawky is the first consumer. You can be the next. Teams plug Feather into their internal agents, in-house performance tooling, or even Claude-based workflows. If you'd like to talk to us about Feather independently of the Performance Agent, we're happy to.

Six categories of living context: Creative Hook Patterns (continuous), Competitor Intel from Meta and Google ad libraries (daily), Trends & Virality across platforms (hourly), Audience Behavior by cohort (real-time), Brand Corpus with voice and category guardrails (weekly), and Performance History covering 18 months of CPA / ROAS / LTV by audience and geo (always-on). Each lives through the four-stage lifecycle: capture → running context → decay & stickiness → propagation.

It complements them. Your warehouse remains the system of record for raw events. Feather is the system of working memory the agent uses to make decisions. We can ingest from Snowflake, BigQuery, Redshift, Databricks, or any reverse-ETL pipeline. Feather doesn't replace your CDP. It sits above it, decaying what doesn't matter and surfacing what does.

Hawky is the agent: the operator that runs your accounts. Feather is the substrate: the memory layer the agent stands on. Hawky pays for Feather. Feather makes Hawky compound. They're a matched pair, but Feather is its own product with its own audience (AI engineers building performance tooling). Don't collapse them. Buying one doesn't lock you out of the other.