Research agents

The shape of a research agent's day

Most research agents have a similar loop: wake up on a schedule, figure out what is new since last run, dive deep on one or two topics, write a synthesis. TensorFeed slots into the "what is new" and "dive deep" steps. Three endpoints cover most of the work.

Step 1: Boot up with the morning brief

One paid call returns a curated 24-hour window:

from tensorfeed import TensorFeed

tf = TensorFeed(token="tf_live_...")

brief = tf.whats_new(days=1, news_limit=10)
# brief["pricing"]["changes"]   - which models had price changes
# brief["pricing"]["new_models"] - any new models launched in the period
# brief["status"]["incidents"]   - any provider outages
# brief["news"]                  - top 10 headlines, newest first

Most days the answer is "nothing dramatic happened." The brief surfaces that cleanly so the agent can skip ahead. On busy days, it is the difference between a one-line decision and a fifteen-call reconciliation.

Step 2: Dive on a specific topic

Two patterns for "dig deeper" research:

By topic / keyword:

# What did Anthropic publish in March?
results = tf.news_search(
    q="Anthropic Claude",
    from_date="2026-03-01",
    to_date="2026-03-31",
    provider="anthropic",
    limit=25,
)
for r in results["results"]:
    print(f"{r['published_at']}: {r['title']}")

By provider (one-call deep-dive):

# Everything about a provider in one paid call
profile = tf.provider_deepdive("anthropic")
# profile["status"]            - live status + components
# profile["models"]            - sorted flagship-first, with benchmark scores joined
# profile["recent_news"]       - top 8 mentions
# profile["agent_traffic_24h"] - hits attributed to Anthropic bots

Step 3: Compare across providers

When the synthesis needs a side-by-side, use the comparison endpoint:

compare = tf.compare_models(ids=[
    "Claude Opus 4.7",
    "GPT-5.5",
    "Gemini 2.5 Pro",
    "DeepSeek V4 Pro",
])
# compare["models"]                          - per-model rows with normalized benchmarks
# compare["rankings"]["cheapest_blended"]    - sorted by blended price
# compare["rankings"]["by_benchmark"]        - per-benchmark leaderboards

Benchmarks are normalized to a union-of-keys with null for missing scores so the agent's downstream code never crashes on undefined.

Step 4: Write the synthesis

That part is your agent's job. TensorFeed gives you the inputs, you decide what to do with them. If the agent runs inside Claude Desktop, the same MCP tools are available natively: "Get the morning brief, then deep-dive on Anthropic, then write a 200-word summary suitable for a daily Slack channel."

Free vs paid

Research agents that want to keep credit usage low can do most of the work on the free tier:

• Free: /api/news, /api/status, /api/models, /api/benchmarks
• Paid: search, deep-dive, compare, whats-new (each 1 credit ~$0.02)

A research agent running once-daily can do its work for under a penny per session using the paid tier; running on free is a full minute of fan-out calls. Pick based on how much your agent cares about latency.

Recommended TensorFeed endpoints (in priority order)

• /api/premium/whats-new: boot-time morning brief
• /api/premium/news/search: full-text search with date/provider filters
• /api/premium/providers/{name}: one-call provider deep-dive
• /api/premium/compare/models: side-by-side comparison with normalized benchmarks
• /api/premium/watches with type:"digest": set-and-forget weekly summaries