AI Agent Routines: Give Recurring Work a Teammate

What is an AI agent routine?

An AI agent routine is a recurring job assigned to a recognizable agent with a defined cadence, expected output, and clear review boundary. The routine determines when the work happens. The agent provides the continuing role, instructions, conversation, and place to report back, while workspace memory provides shared continuity.

The useful formula is:

responsibility + cadence + sources + output + review + owner

A schedule alone is not a routine. "Run every Friday" says when. It does not say what good work looks like, what evidence to use, what to do when nothing changes, or which decision belongs to a person.

For private-capital teams, good routines can keep LP pipelines, investor meetings, research briefs, diligence questions, and relationship follow-through from depending on someone remembering to reopen the process.

The difference between a Skill, a routine, and a teammate

These concepts are related, but they do different jobs.

A Skill can be reused by several people or teammates. A routine is one recurring assignment. A teammate can own several related routines, but should not become a miscellaneous container for every scheduled task.

Key principle:

You define the recurring job. The routine gives it a cadence. The agent gives it an owner.

A routine operating card

Before activating recurring work, write a one-page operating card.

Responsibility

State the outcome the teammate is responsible for preparing. Avoid broad goals such as "manage fundraising." Prefer "keep the active LP pipeline review-ready by identifying material changes, stalled next steps, and missing owners."

Scope

Name the records, stage, geography, fund, strategy, or time window included. Also name exclusions. A first institutional fund and a startup seed round should not share one undifferentiated prospecting routine.

Sources

List the permitted workspace records, connected Apps, documents, and public research sources. Require dated evidence for changing claims. Explain what to do when a source is missing or contradictory.

Cadence

Choose the slowest schedule that still supports the decision. Daily checks are not inherently better than weekly reviews.

Output

Define the artifact and its maximum useful length. Examples include an exception report, a five-prospect update, a meeting brief, a question tracker, or a reviewable draft.

Review boundary

Name the decisions and external actions that require a person. The boundary should be explicit even when no approval is expected on an ordinary run.

No-change behavior

Specify whether a clean run should create a history entry, send a short confirmation, or remain quiet. This prevents the agent from inventing work to look useful.

Stop and pause conditions

Decide when the routine should pause automatically or require a manual review. Examples include a fund close, a disconnected source, repeated errors, an inactive pipeline, or an owner change.

Worked example: weekly LP shortlist review

This example is illustrative, not a customer result.

The assignment

An emerging manager has an active shortlist of 24 prospective LP organizations. Every Friday, the fundraising teammate should review recorded changes and return only the items that may affect next week's work.

The operating card reads:

A useful return

The teammate reports:

Weekly LP review complete. Three items need attention.

1. Northline's next step is nine days overdue. The last recorded commitment was to send the updated portfolio construction note.
2. East Harbor's official site now describes a different minimum allocation. The research source is dated September 25. The fit decision remains unchanged pending review.
3. The Horizon meeting has notes but no owner for the requested legal follow-up.

No messages were sent and no fit decisions were changed.

This return is short because the routine owns exceptions, not a full weekly retelling of the pipeline.

An honest blocked return

If the active-shortlist view is unavailable, the teammate should not fall back to all contacts and pretend the scope is intact. It should say:

Weekly LP review blocked. The active Fund II shortlist could not be loaded. No broader contact search was substituted. Reconnect the view or assign a new source before the next run.

Blocked work is a valid outcome when it prevents unsupported conclusions.

How to set up an AI agent routine

1. Start with work that already repeats

Look for a responsibility someone performs on a recognizable cadence. If the team cannot explain the current process, the agent will inherit that ambiguity.

Good starting jobs include:

• weekly LP pipeline exceptions;
• meeting preparation for the next business day;
• an open-diligence-question review;
• a post-meeting commitment check;
• a monthly relationship-coverage review;
• or a recurring sponsor-investor update preparation check.

2. Assign one accountable teammate

Choose the teammate whose standing responsibility matches the job. Routine ownership should make the system easier to understand. If three agents could plausibly own the same routine, their roles may be too broad.

3. Separate standing instructions from task instructions

Standing instructions explain how the teammate works across assignments. Routine instructions define this job, schedule, source scope, and output.

For example:

• Standing instruction: "Use concise, evidence-led writing. Identify unknowns and separate facts from inferences."
• Routine instruction: "Review active Fund II LP records every Friday and return only material changes, missing owners, and overdue next steps."

This separation makes the role easier to improve without rewriting every schedule.

4. Define the evidence contract

Decide which claims need a source link, observation date, or record reference. Explain how the teammate should handle stale, missing, or conflicting evidence.

5. Define the authority contract

List what the routine can prepare, what it can update, and what it cannot do without confirmation. "Nothing external is sent" is useful only when the described workflow actually enforces it.

6. Test one run before activating the schedule

Use a fixed, representative set of records. Review omissions, unsupported inferences, length, and stop behavior. Correct the instructions before increasing cadence.

7. Review the first several runs

Look for repeated noise, stale sources, silent scope expansion, and decisions that should have stopped for a person. A routine is configured work, not a fire-and-forget employee.

What a run history should tell you

Recurring work needs more than a success timestamp. A useful history answers:

• Why did this run start?
• Which instruction and schedule version applied?
• Which sources were used?
• What output returned?
• What changed?
• Did the routine need review or become blocked?
• Was anything updated or sent?
• Who changed the routine afterward?

This does not require exposing hidden model reasoning. It requires an operational receipt that lets a person understand the work and its consequences.

Google's guidance on long-running agents emphasizes durable state that survives restarts. Anthropic's research on long-running agent harnesses similarly describes the need for explicit progress artifacts across sessions. These sources discuss general agent engineering, not Finta's performance. They support a durable principle: recurring work needs state outside a single transient prompt.

How Aurora agent routines work in Finta

Finta Automations supports scheduled Aurora work. Current schedule options include supported daily, weekly, monthly, annual, or hourly cadences, subject to organization access, credits, connected context, and tool availability.

Aurora Agents give that work a recognizable owner. Each optional teammate has its own instructions, persistent conversation, owned routines, and work history inside the same authorized workspace. Skills, connected Apps, and memory remain workspace-level resources shared with Aurora and every teammate. The teammate can be paused, and its routines can retain visible state rather than silently running as unrelated work.

The global Automations surface and an agent's profile are two views of the same operational responsibility. The important point is not where the edit lives. It is that routine ownership, instructions, schedule, output, and history stay coherent.

The Personal Intelligence Routines guide covers the broader design of responsible recurring work, including scope, permissions, outputs, and stopping conditions. This article owns a narrower question: how to assign that work to a recognizable Aurora teammate and operate it over time.

Common routine failures

Running too often

High cadence can create duplicated work, unnecessary cost, and notification fatigue. Match the schedule to the decision window.

Returning summaries instead of decisions

A five-page recap of the pipeline may be less useful than three evidence-backed exceptions. Design the output around the next review.

Hiding no-change runs

If the team needs proof that a check occurred, record it. If they do not, keep it quiet. Do not confuse silence with failure.

Treating missing data as a negative answer

"No evidence found" is not the same as "this did not happen." The routine should preserve unknown states.

Allowing the role to drift

A routine for pipeline review should not gradually start drafting outreach, changing fit, or making allocation judgments unless the job and authority are deliberately revised.

Losing the human owner

An AI teammate can own preparation and follow-through. A named person still owns the relationship, mandate, and consequential decision.

Start with a routine that makes Friday easier

A practical first routine is usually not the most ambitious. Choose a recurring review that already has a stable source, a known output, and a clear human decision.

Write the operating card. Test one run. Activate the cadence. Review the first returns. Pause it when the job changes.

Use Managing AI Teammates: Status, Activity, and Control for the supervision model. Use the AI Teammates for Private Capital guide to choose the right first responsibility.

Explore Finta Automations to see scheduled Aurora work, editable templates, and reviewable outputs.

Sources and disclosure

• Finta Automations
• Finta Aurora
• Google Developers: Build Long-Running AI Agents That Pause, Resume, and Never Lose Context
• Anthropic: Effective Harnesses for Long-Running Agents
• NIST: AI Agent Standards Initiative

Research updated September 27, 2026. Written by Finta Editorial Team and reviewed by Finta Product and Editorial. The operating card and scenarios are editorial examples, not customer outcomes. Schedule support, connected context, tools, permissions, credits, and feature availability can change. This article provides general operational education, not investment, legal, compliance, privacy, security, broker-dealer, or fundraising-outcome advice.