Field Notes: AI Agents Are Learning to Negotiate Commitments, Not Just Complete Tasks

Most business workflows are described as sequences of tasks: collect the data, draft the proposal, obtain approval, send the document. But the operational risk rarely sits inside the tasks themselves. It sits in the commitments connecting them: who promised what, to whom, by when, under which conditions, and what must happen if the promise can no longer be kept.

AI agents are beginning to operate across those connections. An agent may request information from finance, reserve time on an executive calendar, tell a customer that a revision is coming, and trigger work in another system. Each action creates or modifies an expectation. The agent is no longer merely completing work. It is negotiating a web of commitments.

That distinction matters because task completion is easy to observe, while commitment integrity is not. A polished output can conceal an expired approval, an unmet dependency, or a promise that no authorized person intended to make.

What changed: agents now persist across systems and time

Earlier assistants generally responded within a single interaction. Their work ended when they produced text or returned an answer. Newer agent patterns can maintain state, call tools, wait for external events, resume a workflow, and communicate with multiple parties. This gives them continuity.

Continuity changes the unit of work. Consider an agent asked to coordinate a contract renewal. It may inspect account history, ask legal for revised language, propose meeting times, prepare a pricing exception, and follow up when reviewers are late. The assignment cannot be represented accurately as one task. It contains conditional obligations that mature at different times.

• Temporal obligations: Follow up if legal has not responded by a specified time.
• Conditional obligations: Request executive approval only if the discount exceeds the account team’s authority.
• Reciprocal obligations: Send the revised agreement after the customer supplies required documentation.
• Representational obligations: Do not describe a date, price, or concession as confirmed until an authorized source has committed to it.

The practical shift is from “Can the agent perform this step?” to “Can the agent preserve the status and meaning of every promise while performing it?”

The hidden object is a commitment record

Most workflow systems store tasks, owners, statuses, and due dates. That is not enough for agentic work. A task says what should be done. A commitment record says what expectation has been created and what evidence would satisfy it.

This structure prevents a common failure: treating an outbound message as a harmless communication step. “We will send the final version tomorrow” is not merely text. It creates an expectation, starts a clock, and may expose the business if the underlying approval is incomplete.

What it means in practice: separate preparation from commitment

Agents need different permissions for preparing an action and binding the organization to it. A system may safely allow an agent to draft a refund offer while requiring a manager to approve the amount. It may permit the agent to identify available implementation dates while withholding permission to confirm one with the customer.

A useful operating model has four stages:

1. Propose: The agent constructs a possible commitment and identifies dependencies.
2. Validate: It verifies authority, policy, capacity, and source freshness.
3. Bind: An authorized actor or policy-controlled mechanism activates the commitment.
4. Monitor: The agent tracks fulfillment, detects risk, and initiates recovery.

For example, suppose a customer asks whether onboarding can begin next Monday. The agent finds an open slot on a delivery calendar. That slot is evidence of availability, not evidence of commitment. It must also check whether staffing is confirmed, whether the contract is effective, and whether the person or policy controlling the slot permits reservation. Only then should it bind the date.

This separation may introduce friction. It also prevents agents from converting plausible circumstances into promises. The goal is not universal approval. The goal is binding only when the required predicates are true.

Commitment collisions require explicit arbitration

An agent acting across functions will encounter commitments that cannot all be satisfied. A sales team may promise speed while security requires additional review. A hiring manager may want an interview scheduled while a candidate has restricted availability. A support agent may offer a replacement while inventory is reserved for a recall.

A conventional task manager flags delays. A commitment-aware agent must identify the collision before it communicates certainty. It needs rules for priority and renegotiation, such as:

• Regulatory and safety obligations override commercial preferences.
• Externally communicated commitments take precedence over internal target dates unless a higher-order obligation intervenes.
• Confirmed resource reservations outrank tentative holds.
• When two commitments have equal standing, escalate to the designated accountable owner rather than silently choosing.
• If fulfillment becomes doubtful, renegotiate before the deadline instead of reporting failure afterward.

The agent should also explain the conflict in decision-ready form. “The launch date is at risk” is weak. “The promised launch date depends on security approval; two required controls remain unverified, and the review owner has not accepted the request” identifies the commitment, dependency, and missing action.

Recovery behavior is as important as initial accuracy

No agent can guarantee that every commitment will be fulfilled. People become unavailable, systems fail, suppliers miss dates, and policies change. The more valuable capability is recognizing when the original promise is degrading and managing recovery without inventing reassurance.

Recovery should begin from the recipient’s reliance. If a customer planned a campaign around a delivery date, a silent internal reschedule is not enough. The agent must determine who needs notification, what alternatives are authorized, and whether the revised commitment requires new consent.

A worked recovery example

An operations agent has coordinated a data export for Friday. On Thursday, it detects that a required privacy review remains open. A task-oriented agent might remind the reviewer and continue waiting. A commitment-oriented agent would assess whether Friday remains credible, identify the person authorized to change the date, prepare alternatives, and avoid describing delivery as confirmed.

If policy permits, it could offer a narrower export that already satisfies review requirements. Otherwise, it should escalate with a compact decision: preserve scope and move the date, or preserve the date and reduce scope. The agent is not merely fixing a late task. It is restructuring the commitment while preserving informed consent.

Where current implementations remain weak

Three unresolved problems stand out.

Authority is often implicit

Organizations rarely maintain a machine-readable map of who can promise discounts, delivery dates, policy exceptions, or resource allocations. Authority may depend on geography, account tier, contract type, or current capacity. Without that map, agents either over-escalate or infer authority from unreliable signals such as job titles and previous messages.

Language changes commitment strength

“We are targeting Friday,” “Friday is available,” and “We confirm delivery on Friday” have different operational meanings. Models can recognize these distinctions, but dependable enforcement requires structured states behind the language. Otherwise, cautious wording may conceal a bound action, or confident wording may overstate a tentative plan.

Commitments cross organizational boundaries

One company’s agent cannot assume another company interprets status, approval, or reservation in the same way. Shared protocols may transmit a date or acceptance, yet still omit authority, conditions, and revocation rules. Inter-agent coordination will remain fragile until systems exchange commitment semantics rather than only messages and events.

The operating test for leaders

Before assigning an agent a cross-functional workflow, leaders should inspect every place where the agent can create reliance. Ask whether it can reserve capacity, imply approval, confirm a date, offer a concession, instruct another team, or tell an external party that something will happen.

For each point, define the binding authority, activation conditions, evidence of fulfillment, monitoring interval, and recovery owner. Then test the workflow with a dependency that fails after the agent has begun coordinating. The revealing question is not whether the agent completes the happy path. It is whether it knows which promises are still valid when the path breaks.

The next generation of useful agents will not be distinguished only by better planning or more tool access. They will be distinguished by disciplined commitment management: knowing when an expectation has been created, whether the organization is entitled to create it, and how to repair it before trust is spent.

This post was drafted with AI assistance and reviewed against our editorial policy before publication. Corrections are made at the source, on the page, with the date shown.